BEGIN:VCALENDAR VERSION:2.0 PRODID:-//iisc-demo v1//NONSGML iCal Writer//EN CALSCALE:GREGORIAN METHOD:PUBLISH BEGIN:VEVENT DTEND:20191213T120000Z UID:ab8a27738768baa49648c02855f1c1bb-9 DTSTAMP:19700101T120016Z DESCRIPTION:Cryptography..In Anticipation of Quantum Computer URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/9/cryptography-in-anticipation-of-quantum-computer/ SUMMARY:Does the quantum supremacy has any impact on cryptographic security as practiced today? Will we be prepared enough in case quantum computer becomes a reality in a not-so-distant future? Painting with a broad brush, the talk will attempt to draw the contours of cryptography in a post-quantum world. We will touch upon the challenges and difficulties of securing our various mundane activities that involve communication over a public channel and the kind of research that have been initiated as part of this broad agenda. DTSTART:20191213T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191129T120000Z UID:0f00f61adbb9f6ad45eac8f690077e4d-10 DTSTAMP:19700101T120016Z DESCRIPTION:An Algorithmic View of Smart Living: The Next Frontier URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/10/an-algorithmic-view-of-smart-living-the-next-frontier/ SUMMARY:The advent of wireless sensor networks and Internet of Things (IoTs) are making our lives increasingly dependent on a wide variety of cyber-physical systems (CPS) and smart services (e.g., smart energy, transportation, healthcare, agriculture, etc.), while aiming to improve quality of life. The availability of rich mobile devices like smartphones are also empowering humans to act as sensors to collect fine-grained information via crowd sensing, resulting in actionable inferences and decisions. This talk will first provide a “smart living” vision and highlight the underlying research challenges. Next, it will formulate fundamental problems related to sensor data fusion, coverage and connectivity, mobile charging, security, network lifetime and resource trade-off. Novel solutions will be designed based on randomized and approximation algorithms on graphs, optimization techniques, geometric probability theory, uncertainty reasoning, trust model and game theory. Case studies and experimental results will be presented where possible. The talk will be concluded with emerging applications of sensors and IoTs, followed by directions of future research. DTSTART:20191129T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191206T120000Z UID:9f819b24ca977f0bf6aeb91bea1cb681-12 DTSTAMP:19700101T120011Z DESCRIPTION:Cyber(-Physical) Systems and Machine Learning: Research Challenges URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/12/cyber-physical-systems-and-machine-learning-research-challenges/ SUMMARY:The rapid digital transformation of our world, the exponential improvements in computing, communication, and storage, and the evolution of machine learning are transforming the very nature of scientific discovery and engineering research and making advances in many areas including personalized health care, emergency response, traffic flow management, and electric power generation. These developments have resulted in new complex applications and workflows that require sophisticated cyber and cyber-physical systems that deeply integrate computation, communication, and control into physical systems. After a brief overview of machine learning, I will describe the characteristics of the new applications and the architectures of the required cyber(-physical) systems. The focus will be on the current research challenges in programming, debugging, and managing these systems safely and securely, utilizing machine learning technologies. DTSTART:20191206T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191127T120000Z UID:9a993e6cf62350a4adc08ffc626bdc83-14 DTSTAMP:19700101T120004Z DESCRIPTION:Improved Truthful Mechanisms for Combinatorial Auctions with Submodular Bidders URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/14/improved-truthful-mechanisms-for-combinatorial-auctions-with-submodular-bidders/ SUMMARY:A longstanding open problem in Algorithmic Mechanism Design is to design computationally efficient truthful mechanisms for (approximately) maximizing welfare in combinatorial auctions with submodular bidders. The first such mechanism was obtained by Dobzinski, Nisan, and Schapira [STOC’06] who gave an O(log^2 m)-approximation where m is number of items. This problem has been studied extensively since, culminating in an O(sqrt{log m})-approximation mechanism by Dobzinski [STOC’16]. We present a computationally-efficient truthful mechanism with approximation ratio that improves upon the state-of-the-art by an exponential factor. In particular, our mechanism achieves an O((loglog m)^3)-approximation in expectation, uses only O(n) demand queries, and has universal truthfulness guarantee. This settles an open question of Dobzinski on whether Θ(sqrt{log m}) is the best approximation ratio in this setting in negative. This is based on a joint work with Sepehr Assadi and appeared in FOCS 2019. DTSTART:20191127T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191127T120000Z UID:8e66bd002cea09d869e43b2e174dd475-15 DTSTAMP:19700101T120004Z DESCRIPTION:Beyond trace reconstruction: Population recovery from the deletion channel URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/15/beyond-trace-reconstruction-population-recovery-from-the-deletion-channel/ SUMMARY:Population recovery is the problem of learning an unknown distribution over an unknown set of n-bit strings, given access to independent draws from the distribution that have been independently corrupted according to some noise channel. Recent work has intensively studied such problems both for the bit-flip and erasure noise channels. We initiate the study of population recovery under the deletion channel, in which each bit is independently deleted with some fixed probability and the surviving bits are concatenated and transmitted, in both worst-case and average-case settings of the strings in the support. This is a generalization of trace reconstruction, a challenging problem that has received much recent attention. For the worst case, we show that for any s = o(log n / log log n), a population of s strings from {0,1}^n can be learned under deletion channel noise using exp(n^{1/2+o(1)}) samples. On the lower bound side, we show that n^{Omega(s)} samples are required to perform population recovery under the deletion channel, for all s <= n^0.49. For the average case, we give an efficient algorithm for population recovery. The algorithm runs in time poly(n,s,1/eps) and its sample complexity is poly(s, 1/eps, exp(log^{1/3} n)), where eps is the TV distance between the original and output distributions. This is based on the following joint work with Frank Ban, Xi Chen, Adam Freilich and Rocco Servedio: https://arxiv.org/abs/1904.05532 https://arxiv.org/abs/1907.05964 DTSTART:20191127T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191206T120000Z UID:33848f15b81627c115849298a0c098ad-16 DTSTAMP:19700101T120015Z DESCRIPTION:Geometric Search Techniques for Provably Robust Query Processing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/16/geometric-search-techniques-for-provably-robust-query-processing/ SUMMARY:Relational Database Management Systems (RDBMS) constitute the backbone of today\\\'s information-rich society, providing a congenial environment for handling enterprise data during its entire life cycle of generation, storage, maintenance and processing. The Structured Query Language (SQL) is the de facto standard interface to query the information present in RDBMS based repositories. An extremely attractive feature of SQL is that it is declarative in nature, meaning that the user species only the end objectives, leaving to the system the task of identifying the optimal execution strategy to achieve these objectives. A crucial input to generating efficient query execution strategies, called \\ DTSTART:20191206T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191206T120000Z UID:68dc6eb31cf8d8a32f23568794e6ac89-17 DTSTAMP:19700101T120016Z DESCRIPTION:A Graph-Theorist's Perspective on the Quest for Dichotomy URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/17/a-graph-theorists-perspective-on-the-quest-for-dichotomy/ SUMMARY:The celebrated Feder-Vardi Dichotomy Conjecture for Constraint Satisfaction has recently been established by Bulatov and by Zhuk. Because of the profound impact the conjecture had on theoretical computer science, Feder and Vardi were jointly awarded the Alonzo Church Award for 2019. The solution involved a beautiful blend of graph theory, logic, and universal algebra, developed over a period of 25 years; these developments, in turn, impacted all these fields. I will present a personal account of how a graph-theorist perceived the events and developments leading up to the formulation, and the solution, of the conjecture. I will focus on the impact on graph theory, but briefly mention the other fields as well. DTSTART:20191206T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191219T120000Z UID:d1586b5f05c4f047a2dda3c6f697614f-18 DTSTAMP:19700101T120011Z DESCRIPTION:Discover[i]: Component-based Parameterized Reasoning for Distributed Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/18/discoveri-component-based-parameterized-reasoning-for-distributed-applications/ SUMMARY:This talk begins with an overview of recent and ongoing work in Purdue’s Formal Methods (PurForM) research group. Next, the talk presents our Discover[i] project which seeks to automate reasoning about new classes of distributed applications built on top of verified components. The current focus of this project is on parameterized verification and synthesis of systems that use consensus protocols, such as Paxos, as a building block to provide higher-level functionality. The talk explains the key ingredients of our framework: (1) an abstraction of consensus with a simple atomic primitive, (2) a decidability result and algorithm for parameterized verification of safety properties of systems with such consensus primitives, and (3) an algorithm for parameterized synthesis of coordination for such systems. Discover[i] is joint work with Nouraldin Jaber, Christopher Wagner, Swen Schewe and Milind Kulkarni. DTSTART:20191219T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191211T120000Z UID:4b8373d21d1aed7053f062ae458bd364-19 DTSTAMP:19700101T120010Z DESCRIPTION:Optimistic Hybrid Analysis for System Security and Reliability URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/19/optimistic-hybrid-analysis-for-system-security-and-reliability/ SUMMARY:Dynamic analysis tools such as information-flow tracking (DIFT) and data-race detection are useful for enforcing security policies and improving software reliability. But these tools are rarely used in production systems, as it can slow down a program by an order of magnitude. Static whole program analyses can be used to prove safe execution states and elide unnecessary runtime checks, but in practice, they are mostly ineffective for large programs. The reason is that they are greatly hindered by the need to prove their soundness, as soundness requires analysis of all possible executions and sound over-approximations of a program. This talk presents Optimistic Hybrid Analysis (OHA). OHA improves the scalability and precision of whole program static analysis by one to two orders of magnitude by making optimistic assumptions about a program’s properties that are almost always true, but are hard to prove statically. By making these assumptions, we sacrifice soundness of static analysis, but yet, we preserve soundness of dynamic analysis by checking them at runtime and recovering when they fail. OHA has been used to obtain three promising results. It speeds up FastTrack, a well-known dynamic data-race detector by 3.5x; reduces the overhead of DIFT to less than 10%, a 4.4x improvement; enables the first known solution for a sound garbage collector for C/C++ using efficient pointer provenance. DTSTART:20191211T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191210T120000Z UID:d0557c2448b01b644f8a20d62101eb41-20 DTSTAMP:19700101T120011Z DESCRIPTION:CHET: An Optimizing Compiler for Fully-Homomorphic Neural-Network Inferencing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/20/chet-an-optimizing-compiler-for-fully-homomorphic-neural-network-inferencing/ SUMMARY:Fully Homomorphic Encryption (FHE) refers to a set of encryption schemes that allow computations on encrypted data without requiring a secret key. Recent cryptographic advances have pushed FHE into the realm of practical applications. However, programming these applications remains a huge challenge, as it requires cryptographic domain expertise to ensure correctness, security, and performance. CHET is a domain-specific optimizing compiler designed to make the task of programming FHE applications easier. Motivated by the need to perform neural network inference on encrypted medical and financial data, CHET supports a domain-specific language for specifying tensor circuits. It automates many of the laborious and error prone tasks of encoding such circuits homomorphically, including encryption parameter selection to guarantee security and accuracy of the computation, determining efficient tensor layouts, and performing scheme-specific optimizations. Our evaluation on a collection of popular neural networks shows that CHET generates homomorphic circuits that outperform expert-tuned circuits and makes it easy to switch across different encryption schemes. We demonstrate its scalability by evaluating it on a version of SqueezeNet, which to the best of our knowledge, is the deepest neural network to be evaluated homomorphically. DTSTART:20191210T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191212T120000Z UID:d5472f29f288608af7cab378c0e27a66-22 DTSTAMP:19700101T120015Z DESCRIPTION:Achieving Fairness in the Stochastic Multi-armed Bandit Problem URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/22/achieving-fairness-in-the-stochastic-multi-armed-bandit-problem/ SUMMARY:The classical Stochastic Multi-armed Bandit (MAB) problem provides an abstraction for many real-world decision making problems such as sponsored-search auctions, crowd-sourcing, wireless communication, etc. In this work, we study Fair-MAB, a variant of the MAB problem, where, in addition to the goal of maximizing the sum of expected rewards, the algorithm also has to ensure that each arm is pulled for at least a given fraction of the total number of rounds which imposes an additional fairness constraint on the algorithm. The non-trivial aspect of Fair-MAB arises when the time horizon T is unknown to the algorithm. The treatment of fairness in the MAB problem is either procedural or outcome-based. Procedural notions of fairness require that the decision-making process of the algorithm must satisfy some fairness guarantees. In contrast to this, we consider an outcome-based fairness notion which can be validated based on the outcome of the decisions of the algorithm. Our primary contribution is characterizing a class of Fair-MAB algorithms by two parameters: the unfairness tolerance and the learning algorithm used as a black-box. We define an appropriate notion of fairness and show that our algorithms guarantee fairness independent of the choice of the learning algorithm. We define the notion of fairness-aware regret which naturally extends the conventional notion of regret, and show that the fairness-aware regret of our algorithm matches in order the regret of the black-box learning algorithm in the absence of fairness constraints. Finally, we show via detailed simulations that our algorithm outperforms the best known algorithm for the Fair-MAB problem, both in terms of the regret and in terms of the fairness guarantee that it provides. We also evaluate the cost of fairness in the MAB setting in terms of the conventional notion of regret. DTSTART:20191212T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191216T120000Z UID:4cf56b82e409d7bc1219f457dda6d733-23 DTSTAMP:19700101T120011Z DESCRIPTION:Lessons Developing Conversational AI Interfaces URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/23/lessons-developing-conversational-ai-interfaces/ SUMMARY:This talk will cover building conversational AI using deep learning technologies and lessons learnt in developing conversational interfaces. The first part of the talk will describe recent advances in deep learning that has led to tremendous progress in natural language processing and is making conversational AI a reality. Conversational AI includes intent classification, sequence labeling, understanding dialogs and context, and coming up with responses to users messages. The second part of the talk will address lessons learned developing conversational interfaces. A conversational interface needs to be personable in addressing, adaptive in understanding, and available with many different supported tasks. Overall, key take aways from the talk will be better understanding of (1) deep learning techniques for natural language processing and (2) interaction patterns for building a good conversational interface. DTSTART:20191216T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191217T120000Z UID:c7a0352ec84b478e0eaeffbf6760f9a5-25 DTSTAMP:19700101T120011Z DESCRIPTION:Equivalence test for the trace iterated matrix multiplication polynomial URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/25/equivalence-test-for-the-trace-iterated-matrix-multiplication-polynomial/ SUMMARY:An m-variate polynomial f is affine equivalent to an n-variate polynomial g if m ≥ n and there is a rank n matrix A ∈ Fn×m and b ∈ Fnsuch that f (x) = g(Ax + b). Given blackbox access to f and g (i.e membership query access) the affine equivalence test problem is to determine whether f is affine equivalent to g, and if yes then output a rank n matrix A ∈ Fn×m and b ∈ Fn such that f (x) = g(Ax + b). This problem is at least as hard as graph isomorphism and algebra isomorphism even when the coefficients of f and g are given explicitly (Agarwal and Saxena, STACS 2006), and has been studied in literature by fixing g to be some interesting family of polynomials. In this work, we fix g to be the trace of the product of d, w × w symbolic matrices X1,...,Xd. We call this polynomial Tr-IMMw,d. Kayal, Nair, Saha and Tavenas (CCC 2017) gave an efficient (i.e (mwd) O(1) time)randomized algorithm for the affine equivalence test of the iterated matrix multiplication polynomial IMMw,d, which is the (1,1)-th entry of the product of d w×w symbolic matrices. Although the definitions of Tr-IMMw,d and IMMw,d are closely related and their circuit complexities are very similar, it is not clear whether an efficient affine equivalence test algorithm for IMMw,d implies the same for Tr-IMMw,d. In this thesis, we take a step towards showing that equivalence test for Tr-IMMw,d and IMMw,d have different complexity. We show that equivalence test for Tr-IMMw,d reduces in randomized polynomial time to equivalence test for the determinant (DET), under mild conditions on the underlying field. If the converse is also true then equivalence tests for Tr-IMMw,d and DET are randomized polynomial time equivalent. It would then follow from the work of Gupta, Garg, Kayal and Saha (ICALP 2019) that equivalence test for Tr-IMMw,d over Q is at least as hard as Integer Factoring. This would then be in sharp contrast with the complexity of equivalence test for IMMw,d over Q which can be solved efficiently in randomized polynomial time (by Kayal, Nair, Saha and Tavenas (CCC 2017)). Recent Update: Soon after the thesis is written, we (together with Vineet Nair) have succeeded in showing the converse direction. So, the above conclusion is indeed true! DTSTART:20191217T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191213T120000Z UID:1d956da67b2f79c6b7c652977fc13842-26 DTSTAMP:19700101T120011Z DESCRIPTION:Pushing the Limits of Fairness in Collective Decision-Making URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/26/pushing-the-limits-of-fairness-in-collective-decision-making/ SUMMARY:In this talk, I will discuss fairness in collective decision-making, which includes everyday tasks such as heirs dividing an estate or residents voting over allocation of public budget. First, I will describe our recent work in fair allocation of private goods, where we strengthen individual notions of fairness to groupwise notions. I will also talk about ongoing projects in which we mix ex-ante and ex-post notions of fairness, and aim to achieve the best of both worlds. Finally, I will talk about generalization to allocation of public goods, and how various notions of fairness extend (or not) to this setting. A promised take-away from the talk? Many fascinating open questions and insights into how resolving them can help the society (cue promotion of www.spliddit.org). DTSTART:20191213T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200103T120000Z UID:dbf7b78ee1355ec1ed901e595ea5479f-27 DTSTAMP:19700101T120016Z DESCRIPTION:Ballooning Multi-Armed Bandits URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/27/ballooning-multi-armed-bandits/ SUMMARY:Many common web-based applications such as online q&A forums and online review portals need a scientific way of identifying high quality answers or opinions and distinguishing them from ordinary ones. To tackle this problem, we introduce a new model which we call "ballooning multi-armed bandits" (B-MAB), a novel extension to the classical stochastic MAB model. In the BMAB model, the set of available arms grows (or balloons) over time. In contrast to the classical MAB setting where the regret is computed with respect to the best arm overall, the regret in a BMAB setting is computed with respect to the best available arm at each time. We first observe that the existing stochastic MAB algorithms are not regret-optimal for the B-MAB model. In fact, if the best arm is equally likely to arrive at any time, a sublinear regret cannot be achieved, irrespective of the arrival of other arms. We next present our main result that if the best arm is more likely to arrive in the early rounds, one can achieve sub-linear regret. Making reasonable assumptions on the arrival distribution of the best arm in terms of the thinness of the distribution’s tail, we prove that the proposed algorithm achieves sub-linear, instance-independent regret. We further quantify explicit dependence of regret on the arrival distribution parameters. Application to online Q&A forums, online review platforms, and many other settings is immediate.(Joint work with Ganesh Ghalme, Swapnil Dhamal, Shweta Jain, and Sujit Gujar) DTSTART:20200103T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191220T120000Z UID:51dcf641c60e574d63408826bd879da3-28 DTSTAMP:19700101T120011Z DESCRIPTION:Privacy Preserving Machine Learning via Multi-party Computation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/28/privacy-preserving-machine-learning-via-multi-party-computation/ SUMMARY:Privacy-preserving machine learning (PPML) via Secure Multi-party Computation (MPC) has gained momentum in the recent past. Assuming a minimal network of pair-wise private channels, we propose an efficient four-party PPML framework over rings, FLASH, the first of its kind in the regime of PPML framework, that achieves the strongest security notion of Guaranteed Output Delivery (all parties obtain the output irrespective of adversary\\\'s behaviour). The state of the art ML frameworks such as ABY3 by Mohassel et.al (ACM CCS\\\'18) and SecureNN by Wagh et.al (PETS\\\'19) operate in the setting of 3 parties with one malicious corruption but achieve the weaker security guarantee of abort. We demonstrate PPML with real-time efficiency, using the following custom-made tools that overcome the limitations of the aforementioned state-of-the-art-- (a) dot product, which is independent of the vector size unlike the state-of-the-art ABY3, SecureNN and ASTRA by Chaudhari et.al (ACM CCSW\\\'19), all of which have linear dependence on the vector size. (b) Truncation and MSB Extraction, which are constant round and free of circuits like Parallel Prefix Adder (PPA) and Ripple Carry Adder (RCA), unlike ABY3 which uses these circuits and has round complexity of the order of depth of these circuits. We then exhibit the application of our FLASH framework in the secure server-aided prediction of vital algorithms: Linear Regression, Logistic Regression, Deep Neural Networks, and Binarized Neural Networks. We substantiate our theoretical claims through improvement in benchmarks of the aforementioned algorithms when compared with the current best framework ABY3. All the protocols are implemented over a 64-bit ring in LAN and WAN. Our experiments demonstrate that, for MNIST dataset, the improvement (in terms of throughput) ranges from 24x to 1390x over Local Area Network (LAN) and Wide Area Network (WAN) together. DTSTART:20191220T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191219T120000Z UID:070544a8e679e48adf71def504633649-29 DTSTAMP:19700101T120011Z DESCRIPTION:Memory compression for higher effective capacity and bandwidth URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/29/memory-compression-for-higher-effective-capacity-and-bandwidth/ SUMMARY:Many important client and data center applications need large memory capacity and high memory bandwidth to achieve their performance and energy efficiency goals. Hardware memory compression provides a promising direction to increase effective memory capacity and bandwidth without increasing system cost. This talk focuses on low overhead, hardware-centric compression mechanisms and compressed data management solutions for CPUs and GPUs. DTSTART:20191219T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191220T120000Z UID:db546d68d08c11cbf440da0835d61f99-30 DTSTAMP:19700101T120011Z DESCRIPTION:Language Agnostic Representation Learning for Product Shopping and Discovery URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/30/language-agnostic-representation-learning-for-product-shopping-and-discovery/ SUMMARY:Have you ever wondered how we serve high quality content at Amazon across different languages in product shopping and discovery? In this talk I will touch upon various scientific challenges involved in serving high quality content to our customers. Particularly, when customers are looking for specific products, we select and show items that customers would like to purchase using query-product classification models. Learning a separate model for each language (across different countries Amazon.com vs Amazon.in) is challenging as the amount of (hard) negatives available is sparse and hard to obtain. We solve this problem in a language-agnostic manner using additional structured relationships, such as query-query alignment and product-product alignment. I will discuss how this additional data along with transformer encoder based architecture with scaled self-attention can outperform several state-of-the-art benchmarks. This paper is accepted for WSDM 2020. At the end, I will discuss various job opportunities and academic collaborations to work with us in our brand new division here at Bangalore. DTSTART:20191220T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191227T120000Z UID:9164794be0bda4633e7a5e78c9c28982-31 DTSTAMP:19700101T120011Z DESCRIPTION:Number of near-shortest vectors in lattices URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/31/number-of-near-shortest-vectors-in-lattices/ SUMMARY:For a matrix A, consider the lattice L(A) formed by all integral vectors v in the null-space of A. We ask for which matrices A, the lattice L(A) has only polynomially many near-shortest vectors i.e., vectors whose length is close to the shortest length in L(A). The motivation for this question comes from the fact that we can get a deterministic black-box polynomial identity testing algorithm for any polynomial whose newton polytope has faces described by matrices with the aforementioned property. We show that when the matrix A is totally unimodular (all sub-determinants are 0,+1,or -1) then the lattice L(A) has only polynomially many near-shortest vectors. The proof of this statement goes via a remarkable theorem of Seymour on a decomposition for totally unimodular matrices. The statement generalizes two earlier known results -- the number of near-shortest cycles and the number of near-shorest cuts in a graph are poly-bounded. As a special case, we get PIT for any polynomial of the form det(sum x_i A_i) for rank-1 matrices A_i. DTSTART:20191227T120000Z END:VEVENT BEGIN:VEVENT DTEND:20191227T120000Z UID:b639316acaa5a2434d3d9fb149de00bc-32 DTSTAMP:19700101T120016Z DESCRIPTION:The Bethe and Sinkhorn Permanents of Structured Matrices and its Implications for Profile Maximum Likelihood URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/32/the-bethe-and-sinkhorn-permanents-of-structured-matrices-and-its-implications-for-profile-maximum-likelihood/ SUMMARY:Here we consider the problem of computing the likelihood of the profile of a discrete distribution, i.e. the probability of observing the multiset of element frequencies, and computing a profile maximum likelihood (PML) distribution, i.e. a distribution with the maximum profile likelihood. For each problem we provide polynomial time algorithms that given $n$ i.i.d samples from a discrete distribution, achieve an approximation factor of $expleft(-O(sqrt{n} log n) right)$, improving upon the previous best-known bound achievable in polynomial time of $exp(-O(n^{2/3} log n))$ (Charikar, Shiragur and Sidford, 2019). Through the work of Acharya, Das, Orlitsky and Suresh (2016), this implies a polynomial time universal estimator for symmetric properties of discrete distributions in a broader range of error parameter. We achieve these results by providing new bounds on the quality of approximation of the Bethe and Sinkhorn permanents (Vontobel, 2012 and 2014). We show that each of these are $exp(O(k log(N/k)))$ approximations to the permanent of $N times N$ non-negative matrices with at most $k$ distinct columns, improving upon the previous known bounds of $exp(O(N))$. To obtain our results on PML, we exploit the fact that the PML objective is proportional to the permanent of a certain Vandermonde matrix with $sqrt{n}$ distinct columns. As a by-product of our work we establish a surprising connection between the convex relaxation in prior work (CSS19) and the well-studied Bethe and Sinkhorn approximations. This is in joint work with Moses Charikar and Aaron Sidford. DTSTART:20191227T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200106T120000Z UID:d2189ece54acb242774fcf34f7694b7b-33 DTSTAMP:19700101T120013Z DESCRIPTION:Beyond trace reconstruction: Population recovery from the deletion channel URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/33/beyond-trace-reconstruction-population-recovery-from-the-deletion-channel/ SUMMARY:Population recovery is the problem of learning an unknown distribution over an unknown set of n-bit strings, given access to independent draws from the distribution that have been independently corrupted according to some noise channel. Recent work has intensively studied such problems both for the bit-flip and erasure noise channels. We initiate the study of population recovery under the deletion channel, in which each bit is independently deleted with some fixed probability and the surviving bits are concatenated and transmitted, in both worst-case and average-case settings of the strings in the support. This is a generalization of trace reconstruction, a challenging problem that has received much recent attention. For the worst case, we show that for any s = o(log n / log log n), a population of s strings from {0,1}^n can be learned under deletion channel noise using exp(n^{1/2+o(1)}) samples. On the lower bound side, we show that n^{Omega(s)} samples are required to perform population recovery under the deletion channel, for all s <= n^0.49. For the average case, we give an efficient algorithm for population recovery. The algorithm runs in time poly(n,s,1/eps) and its sample complexity is poly(s, 1/eps, exp(log^{1/3} n)), where eps is the TV distance between the original and output distributions. This is based on the following joint work with Frank Ban, Xi Chen, Adam Freilich and Rocco Servedio: https://arxiv.org/abs/1904.05532 https://arxiv.org/abs/1907.05964 DTSTART:20200106T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200103T120000Z UID:ccead39cd2a64d8eb95da68dce89a090-34 DTSTAMP:19700101T120011Z DESCRIPTION:Architecting Persistent Memory Systems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/34/architecting-persistent-memory-systems/ SUMMARY:Persistent Memory (PM) technologies (also known as Non-Volatile RAM, e.g., Intel’s 3D XPoint) offer the exciting possibility of disk-like durability with the performance of main memory. Persistent memory systems provide applications with direct access to storage media via processor load and store instructions rather than having to rely on performance-sapping software intermediaries like the operating system, aiding the development of high-performance, recoverable software. For example, I envision storage software that provides the safety and correctness of a conventional database management system like PostgreSQL and the performance of an in-memory store like Redis. However, today’s computing systems have been optimized for block storage devices and cannot fully exploit the benefits of PMs. Designing efficient systems for this new storage paradigm requires a careful rethink of computer architectures, programming interfaces, and application software. While maintaining recoverable data structures in main memory is the central appeal of persistent memories, current systems do not provide efficient mechanisms (if any) to do so. Ensuring the recoverability of these data structures requires constraining the order of PM writes, whereas current architectures are designed to reorder memory accesses, transparent to the programmer, for performance. In this talk, I will introduce recently proposed programming interfaces, called persistency models, that will allow programmers to express the required order of PM writes. Then, I will present my work on developing efficient hardware implementations to enforce the PM write order prescribed by persistency models and tailoring software for these new programming interfaces. DTSTART:20200103T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200113T120000Z UID:697c4dfb867d1885a0f4a4c726ee1750-35 DTSTAMP:19700101T120014Z DESCRIPTION:Experiences in Using Reinforcement Learning for Directed Fuzz Testing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/35/experiences-in-using-reinforcement-learning-for-directed-fuzz-testing/ SUMMARY:Directed testing is a technique to analyze user-specified target locations in the program. It reduces the time and effort of developers by excluding irrelevant parts of the program from testing and focusing on reaching the target location. Existing tools for directed testing employ either symbolic execution with heavy-weight program analysis or fuzz testing mixed with fine-tuned heuristics. In this thesis, we explore the feasibility of using a data-driven approach for directed testing. We aim to leverage the data generated by fuzz testing tools. We train an agent on the data collected from the fuzzers to learn the optimal mutation for each program input. The agent then directs the fuzzer towards the target location by instructing the optimal action for each program input. We use reinforcement learning based algorithms to train the agent. We implemented a prototype of our approach and evaluated it on synthetic as well as real-world programs. We also evaluate and compare different reward mechanisms to train the agent. Our evaluation shows that an agent based on reinforcement learning can learn the task for simple programs. However, it is not able to perform better for real-world programs as compared to fuzzers that have no such learning agent. From our experiments, we conclude that data-driven approaches are feasible and should be pursued. Although in the present state, reinforcement learning is not able to compete with state of the art fuzzers, we hope that advancements in reinforcement learning will be able to bridge the gap. DTSTART:20200113T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200109T120000Z UID:d6cd88971a82887c0e0bd3f7482ad62b-36 DTSTAMP:19700101T120011Z DESCRIPTION:Authenticated Encryption URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/36/authenticated-encryption/ SUMMARY:Authenticated encryption (AE) is a symmetric-key cryptographic primitive for providing both confidentiality as well as authenticity. Due to the recent upsurge of communication networks, the era of the so-called Internet of Things, AE is expected to play a key role in securing these networks. Realizing, the importance of AE, several network protocol suites such as TLS, IPSec, SSH, IEEE 802.11i and few others have adopted the current AE standards. Understanding the inefficiencies of the current standards, some standardizing bodies have proposed standardization competitions for new AE submissions; CAESAR and NIST LWC competitions being the two most renowned among them. These competitions have given a huge boost in the domain of AE related researches. The presentation addresses both the directions of AE design and cryptanalysis. 1) It mainly covers the most popular design aspect in AE based research: Lightweight Cryptography, demonstrated with the design of COFB. 2) The talk also elucidates finding weaknesses from an AE scheme e.g, Cryptanalysis, demonstrated with the design Pi-Cipher. The presentation concludes with the summary of the other significant works and ongoing researches along with the possible future research directions. To be specific, the talk shall bring forth the essence of the work done during my research career, highlight my expertise along with my future projections. DTSTART:20200109T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200117T120000Z UID:60cb03329b2eb650e844707de8d4d6f5-37 DTSTAMP:19700101T120016Z DESCRIPTION:Secure Multiparty Computation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/37/secure-multiparty-computation/ SUMMARY:How can sensitive data be processed without introducing a single point of failure? How can several parties perform a joint computation on their secret inputs, say add them up or compute some other statistics, without revealing any additional information to each other except the desired output? Secure multiparty computation is a powerful cryptographic tool for solving this kind of problem. The talk will give an overview of research in the area, covering classical results, connections with other problems, current research directions, and remaining challenges. DTSTART:20200117T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200123T120000Z UID:d9723f04924c6023dd749e4fdf1bf065-38 DTSTAMP:19700101T120016Z DESCRIPTION:Law and Algorithms: A Cryptographic Lens URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/38/law-and-algorithms-a-cryptographic-lens/ SUMMARY:Computer Science has had a complex relationship with Law since its early days. On the one hand, the disciplines are similar in that they both have deep theoretical roots and at the same time are all-encompassing and very applied. On the other hand, one discipline is founded on mathematics, while the other is purely humanistic in nature. Traditionally, the main points of contact between the discipline were centered around intellectual property for algorithms (software and hardware), and regulating the use and sale of products that include encryption algorithms. Recently, however, many more meeting points have emerged, including (but certainly not limited to) regulating the use of statistical risk-assessment and prediction algorithms; Applying traditionally-humanistic concepts such as privacy, bias, transparency, or individuality, to algorithms; Adjudicating and balancing the protection, sharing, and confinement of data; Determining algorithmic intent, awareness, and liability in automated contracts. Furthermore, cryptographic thinking and tools such as computation over encrypted data, multiparty computation, and zero-knowledge proofs emerge as game-changers in various realistic legal scenarios. This talk is a personal account of the emerging landscape of “Law and Algorithms”, shaped by my interactions with Law scholars and fellow Computer Scientists in the past few years. Three classes co-taught to a mixed audience of CS and Law students, where we tried to build bridges between the disciplines, were particularly influential, and also provide starting points for exciting new research. The classes were co-taught with Daniela Caruso, Aloni Cohen, Stacey Dogan, Cynthia Dwork, Ahmed Ghappour, Shafi Goldwasser, Martha Minow, Frank Partnoy, Pat Williams. DTSTART:20200123T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200108T120000Z UID:ad275f6ebb9f16615e209867bc3e904a-39 DTSTAMP:19700101T120016Z DESCRIPTION:MACHINE LEARNING MODELS:FROM BIRTH TO SERVING THE REAL-WORLD URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/39/machine-learning-modelsfrom-birth-to-serving-the-real-world/ SUMMARY:A Machine Learning (ML) Model is born by tting a function around examples sampled from an unknown distribution, and is designed to generalize to other examples drawn from the same distribution. The machine learning edice largely rests on this foundation. However, modern ML models trained for challenging tasks like object detection, speech recognition, and language understanding require huge amounts of labeled data and leave a large carbon footprint. In return, a model once trained needs to serve many diverse real-world settings that do not perfectly match its birth setting. In this talk, I will discuss current research on propping the ML edice against such shifting foundation. We will span over current props like calibration, out-of-sample detection, domain adaptation, domain generalization, and robustness and reect on some futuristic topics. DTSTART:20200108T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200113T120000Z UID:f135d0840365ce46d89525e477c2d18b-40 DTSTAMP:19700101T120014Z DESCRIPTION:Extending program analysis techniques to web applications and distributed systems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/40/extending-program-analysis-techniques-to-web-applications-and-distributed-systems/ SUMMARY:Web-based applications and distributed message-passing systems are two ubiquitous techniques to distribute the execution of an application across multiple machines. Analysis and verification of such programs poses numerous challenges due to the non-linear flow of control and data within the application. In this thesis, we address program analysis problems in the domains of web applications and message-passing systems, and design effective solutions for these problems. Performing end-to-end analysis of web applications using program analysis techniques is a challenge. This is due to multiple reasons, such as client-server interaction, user interaction, and the use of different languages to implement the server-side code and client-side functionality. We propose a technique that translates the entire web application, including server-side and client-side code, into a single-language ``model''. The model is nothing but a standard (i.e., non-web) executable program in the same language as the server-side code. The model is guaranteed to preserve the essential behavioral aspects of the given web application. The upshot is that off-the-shelf tools can be used to analyze the model, without the need to create web-application specific analysis tools. We instantiate our approach in the context of J2EE applications that use JSP to describe pages and Java servlets to implement the server-side code. We have built a tool for the translation of such applications to Java programs. We evaluate our translation tool by converting 5 real world web applications into corresponding Java programs, and, then analyze these programs using three popular third-party program analysis tools - Wala (for static slicing), Java PathFinder (for explicit-state and symbolic model checking), and Zoltar (fault localization by dynamic analysis). With each of these analysis tools we get precise results in most of the cases. Dataflow analysis is a static analysis technique to identify abstract properties of variables at all the points in a program. The second challenge that the thesis addresses is to extend dataflow analysis to distributed asynchronous message-passing systems. Such systems are commonly used to implement distributed mechanisms, protocols, and workflows, in different domains. To obtain good precision, dataflow analysis of message-passing programs needs to somehow skip the traversal of execution paths that read more messages than the number of messages sent so far in the path, as such paths are infeasible at runtime. Current dataflow analysis techniques for message-passing programs either compromise on precision by traversing infeasible paths in addition to feasible paths, or check only simple types of abstract properties (in particular, ones that are representable using finite lattices). This thesis proposes an approach for precise dataflow analysis of message passing asynchronous programs. Our approach traverses only feasible paths, and can check a class of complex abstract properties that need infinite lattice representations. The problem solved by our approach was not known to be decidable so far. Our approach builds on an existing concept in the analysis of parallel systems, namely, coverability, in a novel and involved manner. We have implemented our approach as a tool, and have analyzed its performance on numerous realistic benchmarks. On several of the benchmarks our approach gave more precise results than a baseline analysis that traverses infeasible paths. DTSTART:20200113T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200114T120000Z UID:926b86f9da0ed6660cc5921c621589a1-41 DTSTAMP:19700101T120011Z DESCRIPTION:The Story of Blind Men and the Elephant: Understanding Context Sensitivity URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/41/the-story-of-blind-men-and-the-elephant-understanding-context-sensitivity/ SUMMARY:Context-sensitive methods of program analysis increase the precision of interprocedural analysis by achieving the effect of call inlining. These methods have been defined using different formalisms and hence appear as algorithms that are very different from each other. Some methods define context, whereas some do not. These methods place different kinds of restrictions on the data flow frameworks supported by them and seem to compute different kinds of values. As a consequence, it is difficult to compare the ideas behind these methods in spite of the fact that they solve essentially the same problem. We believe that these incomparable views are similar to blind men views of an elephant called context-sensitivity. We bring out this whole-elephant-view of context sensitivity in program analysis by proposing a unified model of context sensitivity which provides a clean separation between computation of contexts and computation of data flow values. Our model facilitates declarative specifications of context-sensitive methods by capturing the essence of context-sensitivity. We model most of the known context-sensitive methods using our unified model. This modelling uncovers the hidden notion of contexts in some methods, facilitates insightful comparisons between different methods, and facilitates cross fertilization of ideas and suggest interesting improvements in the known methods. Further, our unification also extends every modelled method to bidirectional analyses. This is a joint work with Swati Jaiswal. DTSTART:20200114T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200107T120000Z UID:d537463cf8789cdd737eded1bd6a1f9b-42 DTSTAMP:19700101T120014Z DESCRIPTION:Revisiting Old Abstractions to Design for The Future URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/42/revisiting-old-abstractions-to-design-for-the-future/ SUMMARY:Computing systems are required to achieve continued performance growth with a focus on security and privacy while increasing energy efficiency. To achieve this goal, we need to revisit old abstractions that exist between applications, system software, and hardware. These old abstractions have enabled independent development and optimizations within each layer. However, to achieve performance and energy efficiency required by future applications, computing systems need to unlock new opportunities for optimizations. The approach I am pursuing to unlock these opportunities is creative cross-layer mechanisms. Such mechanisms allow flexibility across various layers to re-design them cohesively and effectively with optimizations not available when these layers were designed in isolation. In this talk, I will use two examples of cross-layer mechanisms in the area of memory virtualization. These two examples will serve as evidence for using such an approach to enable breakthroughs of the future. DTSTART:20200107T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200117T120000Z UID:fcd0d3e94c4554ea84dcd0c79e2604ee-43 DTSTAMP:19700101T120011Z DESCRIPTION:Parameterized Complexity of Network Design Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/43/parameterized-complexity-of-network-design-problems/ SUMMARY:Network Design Problems, which concern designing minimum cost networks that satisfy given set of ``connectivity constrains'', are very well studied in computer science and combinatorial optimization. Almost all these problems are NP-hard, and a number of results are known about them in the realm of approximation algorithms. Parameterized Complexity is a framework for dealing with computational intractability, where (typically) we try to optimally solve those problem instances which admit a ``small cost solution'' or some other nice structural properties. In this talk we will look at some recent results on the parameterized complexity of network design problems, and future directions for research. DTSTART:20200117T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200110T120000Z UID:80551673b2f7b13a32c37e85e5bd6458-44 DTSTAMP:19700101T120011Z DESCRIPTION:Communication Complexity of Byzantine Agreement, Revisited URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/44/communication-complexity-of-byzantine-agreement-revisited/ SUMMARY:As Byzantine Agreement (BA) protocols find application in largescale decentralized cryptocurrencies, an increasingly important problem is to design BA protocols with improved communication complexity. A few existing works have shown how to achieve subquadratic BA under an adaptive adversary. Intriguingly, they all make a common relaxation about the adaptivity of the attacker, that is, if an honest node sends a message and then gets corrupted in some round, the adversary cannot erase the message that was already sent — henceforth we say that such an adversary cannot perform “after-the-fact removal”. By contrast, many (super-)quadratic BA protocols in the literature can tolerate after-the-fact removal. It turns out, as shown in our work, that disallowing after-the-fact removal is necessary for achieving subquadratic-communication BA. In this talk, I will first present a simple quadratic BA protocol. Next, I will show a new subquadratic binary BA construction (of course, assuming no after-the-fact removal) that achieves near-optimal resilience and expected constant rounds under standard cryptographic assumptions and a public-key infrastructure (PKI). In comparison, all known subquadratic protocols make additional strong assumptions such as random oracles or the ability of honest nodes to erase secrets from memory, and even with these strong assumptions, no prior work can achieve the above properties. DTSTART:20200110T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200108T120000Z UID:3cd4ad2ef5106ceee0abd5f654eda958-45 DTSTAMP:19700101T120014Z DESCRIPTION:Scalable and Effective Polyhedral Auto-transformation without using Integer Linear Programming URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/45/scalable-and-effective-polyhedral-auto-transformation-without-using-integer-linear-programming/ SUMMARY:In recent years, polyhedral auto-transformation frameworks have gained significant interest in general-purpose compilation, because of their ability to find and compose complex loop transformations that extract high performance from modern architectures. These frameworks automatically find loop transformations that either enhance locality, parallelism and minimize latency or a combination of these. Recently, focus has also shifting on developing intermediate representations, like MLIR, where complex loop transformations and data-layout optimizations can be incorporated efficiently in a single common infrastructure. Polyhedral auto-transformation frameworks typically rely on complex Integer Linear Programming (ILP) formulations to find affine loop transformations. However, construction and solving these ILP problems is time consuming which increases compilation time significantly. Secondly, loop fusion heuristics in these auto-transformation frameworks are ad hoc, and modeling loop fusion efficiently would further degrade compilation time. In this thesis, we first relax the ILP formulation in the Pluto algorithm. We show that even though LP relaxation reduces the time complexity of the problem, it does not reduce the compilation time significantly because of the complex construction of constraints. We also observe that due to relaxation, sub-optimal loop transformations that result in significant performance degradation may be obtained. Hence, we propose a new polyhedral auto-transformation framework, called Pluto-lp-dfp, that finds efficient affine loop transformations quickly,while relying on Pluto's cost function. The framework decouples auto-transformation into three components: (1) loop fusion and permutation (2) loop scaling and shifting and (3) loop skewing components. In each phase, we solve a Linear Programming (LP) formulation instead of an ILP, thereby resulting in a polynomial time affine transformation algorithm. We propose a data structure, called fusion conflict graph, that allows us to model loop fusion to work in tandem with loop permutations, loop scaling and loop shifting transformations. We describe three greedy fusion heuristics, namely,max-fuse, typed-fuse and hybrid-fuse, of which, the hybrid-fuse and typed-fuse models incorporate parallelism preserving fusion heuristic without significant compilation time overhead. We also provide a characterization of time-iterated stencils that have tile-wise concurrent start and employ a different fusion heuristic in such programs. In our experiments, we demonstrate that Pluto-lp-dfp framework not only finds loop transformations quickly, resulting in significant improvements in compilation time, but also outperforms state-of-the-art polyhedral auto-parallelizers in terms of execution time of the transformed program. We observe that Pluto-lp-dfp is faster than PoCC and Pluto by a geomean factor of 461x and 2.2x in terms of compilation time. On larger NAS benchmarks, Pluto-lp-dfp was faster than Pluto by 246x. PoCC failed to find a transformation in a reasonable amount of time in these cases. In terms of execution time, the hybrid-fuse variant in Pluto-lp-dfp outperforms PoCC by a geomean factor 1.8x, with over 3x improvements in some cases. We also observe that Pluto-lp-dfp is faster than an improved version of Pluto by a factor of 7%, with a maximum performance improvement of 2x. DTSTART:20200108T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200113T120000Z UID:f2e2a1e66576d2296d0ecf7669c6ebb9-46 DTSTAMP:19700101T120016Z DESCRIPTION:Anomaly Detection in Static Networks using Egonets URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/46/anomaly-detection-in-static-networks-using-egonets/ SUMMARY:Anomaly in networks refers to the situation where the networked system, or part of it, shows significant departure from regular or expected behavioral patterns. Anomalies in networks often imply illegal or disruptive activity by the actors in the network. There has been a lot of recent emphasis on developing network monitoring tools that can detect such anomalous activity. Networks can be static, where we have a single snapshot of the system, or dynamic, where we have network snapshots at several points in time. Anomalies can have different meanings in these two scenarios. In static networks, anomaly typically means a local anomaly, in the form of a small anomalous subgraph which is significantly different from the rest of the network. Local anomalies are difficult to detect using simple network-level metrics since the anomalous subnetwork might be too small to cause significant changes to network-level metrics, e.g., network degree. Instead, such anomalies might be detectable if we monitor sub-network level metrics, e.g., degrees of all subgraphs. However, that option is computationally infeasible, as it involves computing total degrees for all O(2^n) subgraphs of an n-node network. We propose a novel anomaly detection method by using egonet p-values, where the egonet of a node is defined as the sub-network spanned by all neighbors of that node. Since there are exactly n egonets, the number of subgraphs being monitored is n, which is a relatively manageable number. We establish theoretical properties of the egonet method. We demonstrate its accuracy from simulation studies involving a broad range of statistical network models. We also illustrate the method on several well-studied network datasets DTSTART:20200113T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200121T120000Z UID:abd927f48b8aa07d502d3ae4ccf18245-47 DTSTAMP:19700101T120010Z DESCRIPTION:Model Extraction defense using Modified Variational Autoencoder URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/47/model-extraction-defense-using-modified-variational-autoencoder/ SUMMARY:Machine Learning as a Service (MLaaS) exposes machine learning (ML) models that are trained on confidential datasets to users in the form of an Application Programming Interface (API). Since the MLaaS models are deployed for commercial purposes the API is available as a pay-per-query service. A malicious user or attacker can exploit these APIs to extract a close approximation of the MLaaS model by training a substitute model using only black-box query access to the API, in a process called model extraction. The attacker is restricted to extract the MLaaS model using a limited query budget because of the paid service. The model extraction attack is invoked by firing queries that belong to a substitute dataset that consists of either (i) Synthetic Non-Problem Domain (SNPD), (ii) Synthetic Problem Domain (SPD), or (iii) Natural Non-Problem Domain (NNPD) dataset. In this work, we propose a novel defense framework against model extraction, using a hybrid anomaly detector composed of an encoder and a detector. In particular we propose a modified Variational Autoencoder, VarDefend, which uses a loss function, specially designed, to separate the encodings of queries fired by malicious users from those by benign users. We consider two scenarios: (i) stateful defense where an MLaaS provider stores the queries made by each client for discovering any malicious pattern, (ii) stateless defense where individual queries are discarded if they are flagged as out-of-distribution. Treating encoded queries from benign users as normal, one can use outlier detection models to identify encoded queries from malicious users in the stateless approach. For the stateful approach, a statistical test known as Maximum Mean Discrepancy (MMD) is used to match the distribution of the encodings of the malicious queries with those of the in-distribution encoded samples. In our experiments, we observed that our stateful defense mechanism can completely block one representative attack for each of the three types of substitute datasets, without raising a single false alarm against queries made by a benign user. The number of queries required to block an attack is much smaller than those required by the current state-of-the-art model extraction defense PRADA. Further, our proposed approach can block NNPD queries that cannot be blocked by PRADA. Our stateless defense mechanism is useful against a group of colluding attackers without significantly impacting benign users. Our experiments demonstrate that, for MNIST and FashionMNIST dataset, proposed stateless defense rejects more than 98% of the queries made by an attacker belonging to either SNPD, SPD or NNPD datasets while rejecting only about 0.05% of all the queries made by a benign user. Our experiments also demonstrate that the proposed stateless approach makes the MLaaS model significantly more robust to adversarial examples crafted using the substitute model by blocking transferability. DTSTART:20200121T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200121T120000Z UID:8d46b894019a2521546dcb6cac7bafde-48 DTSTAMP:19700101T120011Z DESCRIPTION:Novel Neural Architecture for Multi-Hop Question Answering URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/48/novel-neural-architecture-for-multi-hop-question-answering/ SUMMARY:Natural language understanding has been one of the key drivers responsible for advancing the field of AI. To this end, automated Question Answering (QA) has served as an effective way of measuring the language understanding capabilities of AI systems. Our focus in this thesis is on Reading Comprehension style Question Answering (RCQA) task. Reading comprehension is the ability to understand natural language text and answer questions over it. Specifically, we focus on complex questions that require multi-hop reasoning over facts spread across multiple passages. Recently, there has been a surge in the research activities surrounding RCQA task, primarily due to the emergence of large-scale public datasets. For single-hop RCQA datasets, majority of the proposed solutions are based on massively pre-trained Transformer-style models such as BERT. Some of these solutions have exhibited human level performance. Similar solutions have been proposed for the multi-hop RCQA datasets and they have also improved the state-of-the-art. However, we believe that the core challenges involved in the multi-hop RCQA task have not been addressed effectively by existing solutions and hence there is an opportunity to advance the state-of-the-art. We present a novel deep neural architecture, called TAP (Translucent Answer Prediction), to identify answers and evidence (in the form of supporting facts) in an RCQA task requiring multi-hop reasoning. TAP comprises two loosely coupled networks -- Local and Global Interaction eXtractor (LoGIX) and Answer Predictor (AP). LoGIX predicts supporting facts, whereas AP consumes these predicted supporting facts to predict the answer span. The design of LoGIX is inspired by two key design desiderata -- local context and global interaction -- that we identified by analyzing examples of multi-hop RCQA task. The loose coupling between LoGIX and the AP reveals the set of sentences used by the AP in predicting an answer. Therefore, answer predictions of TAP can be interpreted in a translucent manner. We conduct extensive evaluations and analyses on the HotpotQA dataset to understand the characteristics of TAP. TAP achieved state-of-the-art performance on the distractor setting of the HotpotQA dataset. DTSTART:20200121T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200124T120000Z UID:def0137efd255bf0ae1a5b9e2ba43aa9-49 DTSTAMP:19700101T120011Z DESCRIPTION:Fault Aware Read-Copy-Update URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/49/fault-aware-read-copy-update/ SUMMARY:Deferred freeing is the fundamental technique used in Read-Copy-Update (RCU) synchronization technique where reclamation of resources is deferred until the completion of all active RCU read-side critical sections. We observe that faults inside an RCU read-side critical section can indefinitely block writers that are waiting for the completion of RCU readers and also lead to system failures by preventing the reclamation of deferred resources. We show that the impact of such faults in the Linux kernel is global; a fault in one subsystem can propagate and exhaust critical resources in other unrelated subsystems opening a window of opportunity for DoS-based attacks. For example, a fault in a filesystem can exhaust the process ulimit resulting in fork failures. Since, guaranteeing the absence of faults is practically impossible, it is imperative to harden RCU to tolerate faults. We first study the impact of mitigating lockup by termination of the faulty thread, as thread termination is standard approach used by Linux as recovery strategy. We demonstrate the impact of faults in RCU read-side critical sections and present RCU recovery techniques that use novel approaches to detect and isolate effect of such faults. We also discuss system consistency once the fault is handled by our approach. Timely recovery results in a usable system, preserving the user application state and increasing the system's availability. Our evaluation in the Linux kernel shows that our solution can prevent resource exhaustion in the presence of faults with no additional overhead in the absence of faults. DTSTART:20200124T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200122T120000Z UID:c68939cb8d8d418dcddaaae59bc1e85d-50 DTSTAMP:19700101T120010Z DESCRIPTION:Model Extraction and Active Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/50/model-extraction-and-active-learning/ SUMMARY:Machine learning models trained on a confidential dataset are increasingly being deployed for profit. Machine Learning as a Service (MLaaS) has made such models easily accessible to end-users. They can use it directly as a back-box module to query an input sample and get its corresponding prediction. Prior work has shown that it is possible to extract these models. They developed model extraction attacks that extract an approximation of the MLaaS model by making black-box queries to it. However, none of them satisfy all the four criteria essential for practical model extraction: (i) the ability to extract deep learning models, (ii) non-requirement of domain knowledge, (iii) the ability to work with a limited query budget and (iv) non-requirement of annotations. In this work, we propose a novel model extraction framework that makes use of existing active learning techniques and unannotated public data to satisfy all of them. By using only 30% (30,000 samples) of the unannotated public data, our model extraction framework on an average achieves a performance of 4.70x over uniform noise baseline. We further introduce an ensemble active learning technique by combining two existing state-of-the-art active learning techniques, i.e., DeepFool based Active Learning (DFAL) and Coreset active learning. We empirically show that the ensemble active learning technique, in general, performs better than DFAL and it turns out to be a winner in the majority of our experiments. Finally, we show that our proposed model extraction attack cannot be detected by a state-of-the-art detection method, PRADA, that monitors the distribution of distances between queries for deviation from the normal distribution. DTSTART:20200122T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200129T120000Z UID:a8fbefffefe44e0c0f0ec6760b2f6c62-51 DTSTAMP:19700101T120015Z DESCRIPTION:1. The Power of Encounters 2. Algorithmic fairness in online decision-making URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/51/1-the-power-of-encounters-2-algorithmic-fairness-in-online-decision-making/ SUMMARY:1. A secure encounter is an agreement by two anonymous devices to have met at a given time and place. An associated shared secret enables the devices to subsequently confirm their encounter and communicate securely. In this talk, I will sketch how this simple idea enables fascinating new forms of privacy-preserving, contextual, secure communication among personal and IoT devices, and enables users to produce selective evidence of their personhood and physical whereabouts. Encounters enable powerful forms of secure group communication among devices connected by chains of encounters, subject to spatial, temporal, and causality constraints. Applications range from connecting event attendees and virtual guest books to disseminating targeted health risk warnings, soliciting information and witnesses related to an incident, and tracing missing persons, all while maintaining users’ privacy. Encounters also enable selective proofs of device (co-)location at a given time and place. Finally, encounters can provide evidence of a unique physical trajectory, which suggests a human user and promises a new defense to Sybil attacks. 2. There is growing concern about fairness in algorithmic decision making: Are algorithmic decisions treating different groups fairly? How can we make them fairer? What do we even mean by fair? In this talk I will discuss some of our work on this topic, focusing on the setting of online decision making. For instance, a classic result states that given a collection of predictors, one can adaptively combine them to perform nearly as well as the best of those predictors in hindsight (achieve “no regret”) without any stochastic assumptions. Can one extend this guarantee so that if the predictors are themselves fair (according to a given definition), then the overall combination is fair as well (according to the same definition)? I will discuss this and other issues. This is joint work with Suriya Gunasekar, Thodoris Lykouris, and Nati Srebro. DTSTART:20200129T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200123T120000Z UID:e06431579dced24d649ea26440cef524-52 DTSTAMP:19700101T120011Z DESCRIPTION:Rethinking the role of optimization in learning. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/52/rethinking-the-role-of-optimization-in-learning/ SUMMARY:In this talk, I will overview our recent progress towards understanding how we learn large capacity machine learning models. In the modern practice of machine learning, especially deep learning, many successful models have far more trainable parameters compared to the number of training examples. Consequently, the optimization objective for training such models have multiple minimizers that perfectly fit the training data. More problematically, while some of these minimizers generalize well to new examples, most minimizers will simply overfit or memorize the training data and will perform poorly on new examples. In practice though, when such ill-posed objectives are minimized using local search algorithms like (stochastic) gradient descent ((S)GD), the DTSTART:20200123T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200127T120000Z UID:1a3c8eab53b3b6738a2898a341eda57d-53 DTSTAMP:19700101T120009Z DESCRIPTION:Modern Combinatorial Optimization Problems; Balanced Clustering and Fair Knapsack URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/53/modern-combinatorial-optimization-problems-balanced-clustering-and-fair-knapsack/ SUMMARY:In many classical optimization problems, it is desirable to have an output that is equitable in some sense. The property equitability could be defined differently for different optimization problems. We study this property in two classical optimization problems, clustering and knapsack. In the clustering problem, we desire to have a cost of the clustering evenly distributed among the clusters. We study this problem under the name cost-balanced k-clustering. In the knapsack problem, we desire to have a packing which is fair in some sense. We study this problem under the name fair knapsack. In most of the clustering objectives like k-median or k-center, the cost of assigning a client to the cluster is considered to be borne by a client. Algorithms optimizing such objectives might output a solution where few clusters have very large cost and few clusters have a very small cost. Cost-balanced k-clustering problem aims to obtain the clustering which is cost balanced. We consider objective of minimizing maximum cost of each cluster, where the cost of a cluster is sum of distances of all the points in that cluster from the center of that cluster. We define the notion of γ-stability, for γ > 1, for the problem and give a poly time algorithm for 1.5-stable instances of the problem. We give hardness result. We also define the more general version of the problem and name it DTSTART:20200127T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200130T120000Z UID:ec226bfbebf53eb17d6dde8d71ffd1a5-54 DTSTAMP:19700101T120016Z DESCRIPTION:Finding densest sub-graphs without flow computations URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/54/finding-densest-sub-graphs-without-flow-computations/ SUMMARY:Detecting dense components is a major problem in graph mining,and many different notions of a dense subgraph are used in practice. The densest subgraph problem focuses on finding a subgraph with maximum DTSTART:20200130T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200204T120000Z UID:83fa7b74bd8e1a01bc2e8331ee6d68c7-55 DTSTAMP:19700101T120014Z DESCRIPTION:Cryptographic Schemes for Predicate Evaluation and Search on Outsourced Data URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/55/cryptographic-schemes-for-predicate-evaluation-and-search-on-outsourced-data/ SUMMARY:Recent improvements in architecture and networking are fueling ever growing interest in outsourcing data as well as computation over that data. However, this comes at the cost of trusting the service provider with user data as well as computation which often is violated without the user's knowledge. In this work, we propose a few cryptosystems that address various aspects of cryptographic security of outsourced data. Evaluating a predicate function of interest in the encrypted domain is currently one of the challenging questions in this domain. Our first two works in this thesis focus on evaluation of such predicate functions in the encrypted domain. Then we move to the question of performing search on the outsourced cloud data. The question of searching over outsourced data comes in both theoretical and practical flavors. We have pursued both the directions and have come up with solutions for a few selected problems. The first work deals with a simple access control mechanism where a message is encrypted with respect to a set S. A ciphertext can be decrypted with a secret key associated with a set T if and only if T is a subset of S. Such a scheme is called subset predicate encryption (SPE) and can be used as a building block for constructing wildcard IBE (WIBE), wicked IBE (WKD-IBE) as well as CP-ABE for DNF formula evaluation etc. We propose two constructions of SPE in the large-universe setting with constant-size keys. Our first SPE scheme achieves constant-size ciphertext while the second scheme achieves better security assurance namely, adaptive security in the standard model. Chosen Ciphertext Security (CCA) is the strongest notion of security that is often mandatory in practical scenario. As the adversary in this model is active, devising a CCA-secure encryption scheme is often hard. There are a few techniques that convert CPA-secure predicate encryptions (PE) to CCA-security generically but at a non-trivial cost which is typically proportional to the ciphertext size of the underlying CPA-secure scheme. We devise two generic constructions of CCA-secure PEs from CPA-secure pair encoding-based PEs incurring only a small constant amount of additional cost. Our first work on searching in the encrypted domain deals with two related problems of access control-based search. The first problem is to perform keyword search restricted within a privileged set. Available solutions of this problem achieve selective security under parameterized non-standard assumptions. We propose a solution to the so-called Broadcast Encryption with Keyword Search (BEKS) which is efficient and adaptive secure under a standard static assumption. The second problem takes motivation from encrypted email service where a client might search if the newly received mail contains any of the “specified” keywords. We propose a solution to this problem called Key-Aggregate Searchable Encryption (KASE) that allows searching without leaking any new information to the server. As the name suggests, our KASE construction enjoys constant-size secret keys and is proved adaptive secure under a standard static assumption. On the more applied aspects, our next work presents a searchable encryption framework that can support a number of predicate functions simultaneously. To design the framework, we define a few encodings on top of well-studied Hidden Vector Encryption (HVE). Our novelty lies in the optimization of encodings which result in better search complexity for a few queries and at the same time support new types of queries than the state-of-the-art. This framework can be used to compute statistics on encrypted tabular data like relational database management systems. A problem of recent interest is to perform wildcard search in the encrypted domain in the symmetric key setting. Precisely, the search should find the list of keywords w that match with the wildcard queryword w’ where the match denotes wildcard pattern matching. Non-triviality of this problem follows from the unrestricted nature of the wildcards allowed in a queryword. We reduce the problem of wildcard search to that of secure Boolean formula evaluation in the encrypted domain. This results in a protocol called HAWcS, a provably secure construction of wildcard search in the three party setting. HAWcS performs orders of magnitude better than the state-of-the-art as both our theoretical analysis and prototype implementation suggest. The final work improves upon an existing technique of authenticated email search. In this problem, the cloud service provider who stores the outsourced data has to return an efficiently verifiable proof of the search result. The requirement from this protocol is that a thin client should be able to verify the correctness of the search it has requested the server to perform. We propose a more comprehensive solution of authenticated e-mail search in the two party setting. Our prototype implementation gives evidence to the practicability of our proposal. DTSTART:20200204T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200205T120000Z UID:fc0eb270c5ede7eca8c08f6b7c1751e3-56 DTSTAMP:19700101T120016Z DESCRIPTION:Towards Optimal Secure Computation Protocols URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/56/towards-optimal-secure-computation-protocols/ SUMMARY:Secure computation allows a set of mutually distrusting parties to compute a joint function of their private inputs such that the parties only learn the output of the functionality and nothing else about the inputs of the other parties. Secure computation is one of the central primitives in cryptography that encompasses several cryptographic abstractions and has many practical applications. The seminal results from the 1980s showed that every efficiently computable functionality can also be computed securely. However, these protocols were prohibitively inefficient and could only be considered as feasibility results. One of the central problems in cryptography is to construct secure computation protocols that are optimal in all efficiency parameters. In this talk, I will give an overview of my recent works that make progress towards constructing such optimal secure computation protocols. DTSTART:20200205T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200227T120000Z UID:63fa75fed2430786ec881b9e8cdbfdfe-57 DTSTAMP:19700101T120009Z DESCRIPTION:Deep Learning for Bug-Localization and Program Repair URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/57/deep-learning-for-bug-localization-and-program-repair/ SUMMARY:In this thesis, we focus on the problem of program debugging and present novel deep learning techniques for bug-localization and program repair. Deep learning techniques have been successfully applied to a variety of tasks in natural language processing over the years. Although natural languages and programming languages are similar to some extent, the latter have procedural interpretation and richer structure. Applying deep learning techniques to programs presents many novel challenges which arise due to these differences. We address some of these challenges in this thesis. Most of the existing program debugging research is dominated by formal and theory-first approaches. These approaches fail to take advantage of the existing codebases available online in the form of open source software repositories and student assignment submissions to massive open online courses on programming. Recognizing this, researchers have begun to replace expert-designed heuristics with models learned from codebases to improve the performance of the conventional debugging techniques. This thesis shows that it is possible to solve program debugging problems directly from raw programs using deep learning techniques in an end-to-end manner. More specifically, we present three approaches for bug-localization and program repair which are entirely data-driven and learn to perform their task instead of following the steps specified by a domain expert. We first introduce the notion of common programming errors and present a deep neural network based end-to-end technique, called DeepFix, that can fix multiple such errors in a program without relying on any external tool to locate or fix them. At the heart of DeepFix is a multi-layered sequence-to-sequence neural network with attention mechanism, comprising an encoder recurrent neural network (RNN) to process the input and a decoder RNN with an attention mechanism that generates the output. The network is trained on a labeled dataset to predict a faulty program location along with the correct program statement. Multiple errors in a program can be fixed by invoking DeepFix iteratively. Our experiments demonstrate that DeepFix is effective and fixes thousands of programs. While repositories containing erroneous programs are easily available, the labels of these programs (faulty program location and correct statement) required by DeepFix are not easily available. Labeling a large number of erroneous programs is a daunting task. To address this issue, we propose a novel deep reinforcement learning based technique, called RLAssist, that does not require labeled training data and still matches the performance of DeepFix. At the core of RLAssist is a novel programming language correction framework amenable to reinforcement learning. The framework allows an agent to mimic human actions for text navigation and editing. We demonstrate that the agent can be trained through self-exploration directly from the raw input, that is, the program text itself, without any prior knowledge of the formal syntax of the programming language. Reinforcement learning techniques are, however, usually slow to train. We also show that RLAssist can be trained much faster with the help of expert demonstrations for as little as one-tenth of its training data, which also helps it in achieving better performance than DeepFix. Finally, we present a deep learning based technique for semantic bug-localization in programs with respect to failing tests. The proposed technique works in two phases. In the first phase, a novel tree convolutional neural network is used to predict whether a program passes or fails the given test. In the second phase, we query a state-of-the-art neural prediction attribution technique to find out which lines of the programs make the network predict the failures to localize the bugs. This is a static bug-localization technique and does not require program instrumentation and multiple executions necessary for the existing dynamic bug-localization techniques. Our experiments show that the proposed technique is competitive with two state-of-the-art program-spectrum based and one syntactic difference based bug-localization baselines. All the techniques proposed in this thesis are programming language agnostic. We believe that the ideas and tools developed in this work can potentially be a road-map for future attempts at applying deep learning techniques to more problems in software engineering research. DTSTART:20200227T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200226T120000Z UID:f018461fc66651f8fce4e1e2e24c3923-59 DTSTAMP:19700101T120011Z DESCRIPTION:Security and Privacy of Connected Autonomous Vehicles URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/59/security-and-privacy-of-connected-autonomous-vehicles/ SUMMARY:The upcoming smart transportation systems which consist of connected autonomous vehicles, are poised to transform our everyday life. The sustainability and growth of these systems to their full potential will significantly depend on the robustness of these systems against security and privacy threats. Unfortunately, the communication protocols employed in these systems lack mainstream network security capabilities due to energy constraints of the deployed platforms and bandwidth constraints of the communication medium. In this talk, I will present the results of my efforts in anatomizing the two vital communication protocols employed in the smart transportation: (1) vehicle-to-everything (V2X) communication protocol which is utilized to facilitate wireless communication among connected vehicles, and (2) controller area network (CAN) protocol which is utilized within an autonomous vehicle to enable real-time control of critical automotive components including brakes. For each of these two protocols, I will first describe the inquisitive approach which led to the discovery of the new security vulnerabilities. Then, through the experiments on real-world systems, I will demonstrate how these vulnerabilities can be exploited to launch malicious attacks which evade the state-of-the-art defense mechanisms employed in these systems. I will conclude the talk by discussing novel countermeasures which are required to mitigate these fundamental vulnerabilities and prevent their exploitation. DTSTART:20200226T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200228T120000Z UID:6b075f4ae6797a543902ff072c74daf5-60 DTSTAMP:19700101T120011Z DESCRIPTION:Towards automated debugging and optimization URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/60/towards-automated-debugging-and-optimization/ SUMMARY:Debugging and optimization are largely ad-hoc manual processes taking up 35-75 percent of programmers' time costing more than $100B annually. This process becomes further exacerbated in the modern programming paradigm where programmers stand on the DTSTART:20200228T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200218T120000Z UID:a2289fd2736172a68708e736322fddc8-61 DTSTAMP:19700101T120011Z DESCRIPTION:Efficient Distance Approximation for Structured High-Dimensional Distributions via Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/61/efficient-distance-approximation-for-structured-high-dimensional-distributions-via-learning/ SUMMARY:We design efficient distance approximation algorithms for several classes of structured high-dimensional distributions. Specifically, we show algorithms for the following problems: – Given sample access to two Bayes networks P1 and P2 over known directed acyclic graphs G1 and G2 having n nodes and bounded in-degree, approximate dTV (P1, P2) to within additive error ε using poly(n, ε) samples and time – Given sample access to two ferromagnetic Ising models P1 and P2 on n variables with bounded width, approximate dTV (P1, P2) to within additive error ε using poly(n, ε) samples and time – Given sample access to two n-dimensional gaussians P1 and P2, approximate dTV (P1, P2) to within additive error ε using poly(n, ε) samples and time – Given access to observations from two causal models P and Q on n variables that are defined over known causal graphs, approximate dTV (Pa, Qa) to within additive error ε using poly(n, ε) samples, where Pa and Qa are the interventional distributions obtained by the intervention do(A = a) on P and Q respectively for a particular variable A. Our results are the first efficient distance approximation algorithms for these well-studied problems. They are derived using a simple and general connection to distribution learning algorithms. The distance approximation algorithms imply new efficient algorithms for tolerant testing of closeness of the above-mentioned structured high-dimensional distributions. (based on a joint work with Sutanu Gayen, Kuldeep S. Meel and N. V. Vinodchandran) DTSTART:20200218T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200224T120000Z UID:03d5dc5f14469db0a0d01b071b65362b-62 DTSTAMP:19700101T120016Z DESCRIPTION:Stochastic Approximation in Optimization, Estimation and Reinforcement Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/62/stochastic-approximation-in-optimization-estimation-and-reinforcement-learning/ SUMMARY:​Stochastic approximation deals with the problem of finding zeros of a function expressed as an expectation of a random variable. In this thesis we study and analyze convergence of stochastic approximation algorithms in the context of optimization under uncertainty, statistical estimation and in the context of learning, in particular reinforcement learning. Moreover we also explore second order methods in the context of Reinforcement learning algorithms. Stochastic optimization (SO) considers the problem of optimizing an objective function in the presence of noise. Most of the solution techniques in SO estimate gradients from the noise corrupted observations of the objective and adjust parameters of the objective along the direction of t​he estimated gradients to obtain locally optimal solutions. Two prominent algorithms in SO namely Random Direction Kiefer-Wolfowitz (RDKW) and Simultaneous Perturbation Stochastic Approximation (SPSA) obtain noisy gradient estimate by randomly perturbing all the parameters simultaneously. This forces the search direction to be random in these algorithms and causes them to suffer additional noise on top of the noise incurred from the samples of the objective. Owing to this additional noise, the idea of using deterministic perturbations instead of random perturbations for gradient estimation has also been studied. Two specific constructions of the deterministic perturbation sequence using lexicographical ordering and Hadamard matrices have been explored and encouraging results have been reported in the literature. In this thesis, we characterise the class of deterministic perturbation sequences that can be utilised in the RDKW algorithm. This class expands the set of known deterministic perturbation sequences available in the literature. Using our characterization we propose a construction of a deterministic perturbation sequence that has the least possible cycle length among all deterministic perturbations. Through simulations we illustrate the performance gain of the proposed deterministic perturbation sequence in the RDKW algorithm over the Hadamard and the random perturbation counterparts. We establish the convergence of the RDKW algorithm for the generalized class of deterministic perturbations utilizing stochastic approximation techniques. In statistical estimation one of the popular measures of central tendency that provides better representation and interesting insights of the data compared to the other measures like mean and median is the metric mode. If the analytical form of the density function is known, mode is an argument of the maximum value of the density function and one can apply optimization techniques to find the mode. In many of the practical applications, the analytical form of the density is not known and only the samples from the distribution are available. Most of the techniques proposed in the literature for estimating the mode from the samples assume that all the samples are available beforehand. Moreover, some of the techniques employ computationally expensive operations like sorting. In this thesis we provide a computationally effective, on-line iterative algorithm that estimates the mode of a unimodal smooth density given only the samples generated from the density. Asymptotic convergence of the proposed algorithm using stochastic approximation techniques is provided. We also prove the stability of the mode estimates by utilizing the concept of regularisation. Experimental results further demonstrate the effectiveness of the proposed algorithm. In a discounted reward Markov Decision Process (MDP), the objective is to find the optimal value function, i.e., the value function corresponding to an optimal policy. This problem reduces to solving a functional equation known as the Bellman equation and a fixed point iteration scheme known as the value iteration is utilized to obtain the solution. In literature, a successive over-relaxation based value iteration scheme is proposed to speed-up the computation of the optimal value function. The speed-up is achieved by constructing a modified Bellman equation that ensures faster convergence to the optimal value function. However, in many practical applications, the model information is not known and we resort to Reinforcement Learning (RL) algorithms to obtain optimal policy and value function. One such popular algorithm is Q-learning. In this paper, we propose Successive Over-Relaxation (SOR) Q-learning. We first derive a modified fixed point iteration for SOR Q-values and utilize stochastic approximation to derive a learning algorithm to compute the optimal value function and an optimal policy. We then prove the almost sure convergence of the SOR Q-learning to SOR Q-values. Finally, through numerical experiments, we show that SOR Q-learning is faster compared to the standard Q-learning algorithm. Value iteration is a fixed point iteration technique utilized to obtain the optimal value function and policy in a discounted reward Markov Decision Process (MDP). Here, a contraction operator is constructed and applied repeatedly to arrive at the optimal solution. Value iteration is a first order method and therefore it may take a large number of iterations to converge to the optimal solution. Successive relaxation is a popular technique that can be applied to solve a fixed point equation. It has been shown in the literature that, under a special structure of the MDP, successive over-relaxation technique computes the optimal value function faster than standard value iteration. In this work, we propose a second order value iteration procedure that is obtained by applying the Newton-Raphson method to the successive relaxation value iteration scheme. We prove the global convergence of our algorithm to the optimal solution asymptotically and show the second order convergence. Through experiments, we demonstrate the effectiveness of our proposed appro​ach. DTSTART:20200224T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200226T120000Z UID:fb4cc9ddf1731cfc0d79e021518feb49-63 DTSTAMP:19700101T120010Z DESCRIPTION:On the Round Complexity Landscape of Secure Multi-party Computation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/63/on-the-round-complexity-landscape-of-secure-multi-party-computation/ SUMMARY:In secure multi-party computation (MPC), n parties wish to jointly perform a computation on their private inputs in a secure way, so that no adversary A corrupting a coalition of t among the n parties can learn more information than their outputs (privacy), nor can they affect the outputs of the computation other than by choosing their own inputs (correctness). The round complexity of MPC protocols is a fundamental question in the area of secure computation and its study constitutes a phenomenal body of work in the MPC literature. The research goal of this thesis is to advance the state of the art by expanding this study of round complexity to various adversarial settings and network models. The questions addressed in the thesis are of both theoretical and practical importance. In this talk, we first present a high-level overview of our results which involves establishing new lower bounds on the round complexity of MPC under various settings and constructing upper bounds with optimal round complexity. We then elaborate on one of our recent results that focuses on the exact round complexity of fair and robust MPC protocols in a generalized adversarial setting. Fairness (adversary obtains the output iff honest parties do) and Robustness (adversary cannot prevent honest parties from obtaining the output) are two of the most sought-after properties of MPC protocols. Achieving both, however, brings in the necessary requirement that only upto minority of the parties can be actively corrupt (where actively corrupt parties are allowed to deviate from the protocol in any arbitrary manner). In a generalized adversarial setting where the adversary is allowed to corrupt both actively and passively (weaker adversarial model where corrupt parties follow protocol specifications but try to learn more information than they are supposed to know), the necessary bound for a n-party fair or robust protocol turns out to be t_a + t_p < n, where t_a, t_p denote the threshold for active and passive corruption with the latter subsuming the former. Subsuming active minority as a boundary special case, this setting, denoted as dynamic corruption, opens up a range of possible corruption scenarios for the adversary. While dynamic corruption includes the entire range of thresholds for (t_a, t_p) starting from (ceil(n/2) – 1, floor(n/2)) to (0, n − 1), the boundary corruption restricts the adversary only to the boundary cases of (ceil(n/2) – 1, floor(n/2)) and (0, n − 1). Notably, both corruption settings empower an adversary to control majority of the parties, yet ensuring the count on active corruption never goes beyond ceil(n/2) – 1. We target the round complexity of fair and robust MPC tolerating dynamic and boundary adversaries. References: [1] Arpita Patra, Divya Ravi. On the Power of Hybrid Networks in Multi-Party Computation. IEEE Transactions on Information Theory 2018. [2] Arpita Patra, Divya Ravi. On the Exact Round Complexity of Secure Three-Party Computation. CRYPTO 2018. [3] Megha Byali, Arun Joseph, Arpita Patra, Divya Ravi. Fast Secure Computation for Small Population over the Internet. ACM CCS 2018. [4] Arpita Patra, Divya Ravi, Swati Singla. On the Exact Round Complexity of Best-of-both-Worlds Multi-party Computation. Under Submission. [5] Arpita Patra, Divya Ravi. Beyond Honest Majority: The Round Complexity of Fair and Robust Multi-party Computation. ASIACRYPT 2019. DTSTART:20200226T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200305T120000Z UID:7148e8852c2a2a0a30e0313cc1413981-64 DTSTAMP:19700101T120011Z DESCRIPTION:Fairness in Algorithmic Decision Making URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/64/fairness-in-algorithmic-decision-making/ SUMMARY:Algorithmic (data-driven) decision making is increasingly being used to assist or replace human decision making in domains with high societal impact, such as banking (estimating creditworthiness), recruiting (ranking job applicants), judiciary (offender profiling), healthcare (identifying high-risk patients who need additional care) and journalism (recommending news-stories). Consequently, in recent times, multiple research works have uncovered the potential for bias (unfairness) in algorithmic decisions in different contexts, and proposed mechanisms to control (mitigate) such biases. However, the emphasis of existing works has largely been on fairness in supervised classification or regression tasks, and fairness issues in other scenarios remain relatively unexplored. In this talk, I will cover our recent works on incorporating fairness in recommendation and matching algorithms in multi-sided platforms, where the algorithms need to fairly consider the preferences of multiple stakeholders. I will discuss the notions of fairness in these contexts and propose techniques to achieve them. I will conclude the talk with a list of open questions and directions for future work. DTSTART:20200305T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200228T120000Z UID:ac5b66a404f0563b62cdfba2b6111663-66 DTSTAMP:19700101T120016Z DESCRIPTION:Zero-cost synchronization: Fact or fiction? RCU and other stories URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/66/zero-cost-synchronization-fact-or-fiction-rcu-and-other-stories/ SUMMARY:With proliferation of commodity multicore systems, low cost synchronization is critical for scaling. The read-copy-update (RCU) mechanism has been used for this purpose since the last 4 decades. We analyse this and a few other related mechanisms, and discuss our work in this context (such as memory management, virtualization and fault tolerance) DTSTART:20200228T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200304T120000Z UID:1cac2b798b4805da4511a453c11a606b-67 DTSTAMP:19700101T120009Z DESCRIPTION:On Learning and Lower Bound Problems Related to Iterated Matrix Multiplication Polynomial URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/67/on-learning-and-lower-bound-problems-related-to-iterated-matrix-multiplication-polynomial/ SUMMARY:The iterated matrix multiplication polynomial (IMM) of width w and length d is the 1x1 entry in the product of d square matrices of size w. The w^2d entries in the d matrices are distinct variables. In this thesis, we study certain learning and lower bound problems related to IMM. Our first work gives a polynomial time randomized algorithm for equivalence testing of IMM. At its core, the equivalence testing algorithm exploits a connection between the irreducible invariant subspaces of the Lie algebra of the group of symmetries of a polynomial f that is equivalent to IMM and the layer spaces of a full-rank algebraic branching program computing f. This connection also helps determine the group of symmetries of IMM and show that IMM is characterized by its group of symmetries. Our second work is related to learning affine projections of IMM, which is believed to be a very hard problem as it is equivalent to reconstructing a powerful model to compute polynomials called algebraic branching programs (ABP). Equivalence test for IMM can be viewed as reconstructing ABPs in the average-case, when the width of the ABP is at most (n/d)^0.5, where n and d are the number of variables and the degree of the polynomial computed by the ABP respectively. Our second work improves this by first considering a related problem called `linear matrix factorization’ (LMF) which is a natural generalization of the polynomial factorization problem. We give a polynomial time randomized algorithm for average-case LMF for matrix products of width at most 0.5(n^0.5). In fact, we give a polynomial time randomized algorithm that solves (worst-case) LMF problem when the input matrix product is non-degenerate or pure product-- a notion we define in this work. Using our algorithm for LMF, we give a non-trivial average-case reconstruction algorithm for ABPs of width at most 0.5(n^0.5), which is interesting in the context of the Ω(n^0.5) width lower bound known for homogeneous ABPs. Our last work gives lower bounds on interesting restrictions of arithmetic formulas computing IMM. We prove a w^Ω(d) lower bound on the size of multilinear depth three formulas computing IMM of width w and length d. The lower bound is proved by introducing a novel variant of the partial derivatives measure called skewed partial derivatives, which found applications in other subsequent works. Improving this result to w^Ω(log d) size lower bound on multilinear formulas computing IMM would imply a super-polynomial separation between ABPs and arithmetic formulas. We also show an exponential separation between multilinear depth three and multilinear depth four formulas which was an improvement over the quasi-polynomial separation already known. We also consider a restriction of multilinear formulas, called interval set-multilinear formulas computing IMM. Proving a super-polynomial size lower bound on interval set-multilinear formulas computing IMM would imply a super-polynomial separation between algebraic branching programs and homogeneous formulas in the non-commutative world. We make progress in this direction by giving a super-polynomial size lower bound on an interesting restriction of the interval set-multilinear formula computing IMM.. DTSTART:20200304T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200313T120000Z UID:ba5f7f43f40a86df3cad656e2a54c8ee-68 DTSTAMP:19700101T120013Z DESCRIPTION:Fair Rent Division URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/68/fair-rent-division/ SUMMARY:The theory of fair division addresses the fundamental problem of allocating resources among agents with equal entitlements but distinct preferences. In this talk, I will focus on the classic problem of fair rent division that entails splitting the rent (money) and allocating the rooms (indivisible resources) of an apartment among roommates (agents) in a fair manner. Here, a distribution of the rent and an accompanying allocation is said to be fair if it is envy free, i.e., under the imposed rents, no agent has a strictly stronger preference (utility) for any other agent’s room. While envy-free solutions are guaranteed to exist under reasonably general utility functions, efficient algorithms for finding them were known only for quasilinear utilities. Our work addresses this notable gap and develops approximation algorithms for fair rent division with minimal assumptions on the utility functions. Specifically, we show that if the agents have continuous, monotone decreasing, and piecewise- linear utilities, then the fair rent division problem admits a fully polynomial-time approximation scheme (FPTAS). We complement the algorithmic results by proving that the fair rent division problem (under continuous, monotone decreasing, and piecewise-linear utilities) lies in the intersection of the complexity classes PPAD and PLS. This talk is based on a joint work with Eshwar Arunachaleswaran and Siddharth Barman that appeared in SODA 2019. DTSTART:20200313T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200305T120000Z UID:ba7fa7f71c9a86a4fac996f3ddd0ae0a-69 DTSTAMP:19700101T120016Z DESCRIPTION:Fun facts on compiler optimization and reliability URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/69/fun-facts-on-compiler-optimization-and-reliability/ SUMMARY:If you write an expression in your C program like 'a+b', its evaluation can proceed in either order (e.g., either first 'a' then 'b', or first 'b' then 'a'). I will show how this can be exploited by the compiler to obtain performance improvements of up to 2.6x on real benchmarks, and up to 18x on micro-benchmarks. I will also show that coding patterns that allow such non-determinism to be leveraged for optimization, are very common in real-world code. This part of the talk is based on our PLDI 2020 paper. I will also discuss the astounding complexity of today's optimizing compilers and why you should not trust them. I will then present our recent work on scalable translation validation, that can verify the result of a compilation, and rebuild this lost trust in modern optimizing compilers. This part of the talk is based on our APLAS17, HVC17, SAT18 papers and our ongoing work. The contents of this talk are based on the work done by members of the superoptimizer research group at IIT Delhi. DTSTART:20200305T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200305T120000Z UID:b4279e9b867f97d9b6156ecc94b49ff8-70 DTSTAMP:19700101T120014Z DESCRIPTION:Optimizing the Linear Fascicle Evaluation Algorithm for Multi-Core and Many-Core Systems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/70/optimizing-the-linear-fascicle-evaluation-algorithm-for-multi-core-and-many-core-systems/ SUMMARY:Sparse matrix-vector multiplication (SpMV) operations are commonly used in various scientific and engineering applications. The performance of the SpMV operation often depends on exploiting regularity patterns in the matrix. Various new representations and optimization techniques have been proposed to minimize the memory bandwidth bottleneck arising from the irregular memory access pattern. Among recent representation techniques, tensor decomposition is a popular one used for very large but sparse matrices. Post sparse-tensor decomposition, the new representation involves indirect accesses, making it challenging to optimize for multi-core architectures and even more demanding for the massively parallel architectures, such as on GPUs. Computational neuroscience algorithms often involve sparse datasets while still performing compute-intensive operations. The Linear Fascicle Evaluation (LiFE) application is a popular neuroscience algorithm used for pruning brain connectivity graphs. The datasets employed herein involve the Sparse Tucker Decomposition (STD) - a widely used tensor decomposition method. Using this decomposition leads to multiple irregular array references, making it very difficult to optimize for multi-cores and GPUs. Recent implementations of the LiFE algorithm show that its SpMV operations are the key bottleneck for performance and scaling. In this work, we first propose target-independent techniques such as (1) data restructuring techniques to minimize the effects of irregular accesses, and (2) simple compiler optimizations. Then we apply target-specific optimizations to exploit the resources provided by the architecture. The CPU-specific optimizations that we incorporated are loop tiling, loop parallelization and utilizing BLAS calls to exploit data reuse, coarse-grained parallelism and fine-grained parallelism respectively. We extend the PolyMage domain-specific language, embedded in Python, to automate the CPU-based optimizations developed for this algorithm. Next, we propose various GPU-specific optimizations to optimally map threads at the granularity of warps, thread blocks and grid, and methods to split the computation among thread blocks to obtain fine-grained parallelism and data reuse. Our highly optimized and parallelized CPU implementation obtain a reduction in execution time from 225 min to 8.2 min over the original sequential approach running on 16-core Intel Xeon Silver (Skylake-based) system. Our optimized GPU implementation achieves a speedup of 5.2x over a reference optimized GPU code version on NVIDIA's GeForce RTX 2080 Ti GPU, and a speedup of 9.7x over our highly optimized and parallelized CPU implementation. DTSTART:20200305T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200501T120000Z UID:2e2d121151847cc4a07264954e32a53f-71 DTSTAMP:19700101T120012Z DESCRIPTION:Algorithms for Fair Decision Making: Provable Guarantees and Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/71/algorithms-for-fair-decision-making-provable-guarantees-and-applications/ SUMMARY:The topic of fair allocation of indivisible items has received significant attention because of its applicability in several real-world settings. This has led to a vast body of work focusing on defining appropriate fairness notions, providing existence guarantees, and handling computational issues in obtaining such fair allocations. This thesis addresses open questions in the space of fair allocation of indivisible items. We study constrained versions of the fair allocation problem, namely cardinality constraints and matroid constraints. These constrained settings generalize the unconstrained formulation and we are the first to study these constructs. We establish the existence of well-studied fairness notions (such as EF1 and MMS) under the constrained settings, and we design methods that provide an algorithmic anchor to these existence results. Moreover, we define strictly stronger notions of fairness and provide algorithms for obtaining these stronger fairness guarantees. Finally, we investigate fairness in diverse application scenarios, such as recommendation systems and classification problems. The key novelty involves providing solutions to such applications through the lens of fair allocation. <br><br> <b>Fair Allocation under Cardinality Constraints</b> <br> We investigate the problem of fairly allocating goods under cardinality constraints and additive valuations. In this setting, the set of goods are categorized, and an upper limit is imposed on the number of goods allocated to any agent from a particular category. The objective is to find an allocation that satisfies the given cardinality constraints as well as a fairness constraint. We design an efficient algorithm that computes an envy-free up to one good (EF1) allocation. Additionally, this algorithm outputs an exact maximin share (MMS) allocation when the valuations are binary. We also show that the constrained fair allocation problem with additive valuations reduces to an unconstrained fair allocation problem with submodular valuations. This allows us to guarantee 1/3-approximate maximin share (1/3-MMS) allocations under cardinality constraints. <br><br> <b>Fair Allocation under Matroid Constraints</b> <br>We consider the fair allocation problem under more general constraints. Here, each allocated bundle, in addition to the fairness criterion, needs to satisfy the independence criterion specified by a matroid. We establish that both EF1 and MMS exist in this setting when the valuations are identical. We provide an algorithm that efficiently computes an EF1 allocation. The algorithm initializes an allocation by computing a matroid feasible partition of goods (using a method proposed by Gabow and Westermann) and then iteratively reallocates goods between the bundles till an EF1 allocation is obtained. Our reallocation strategy maintains matroid feasibility at each iteration (using an extension of the strong basis exchange lemma) and also ensures polynomial time convergence. <br><br> <b>Stronger Notions of Fairness </b> <br>We define two novel fairness notions, namely envy-free up to one less preferred good (EFL) and groupwise maximin share (GMMS). We show that these fairness notions are better, in terms of social welfare, compared to EF1 and MMS, respectively. We provide a scale of fairness to establish how these new fairness notions fit in the hierarchy of existing notions. We provide an algorithm that outputs an EFL and ½-GMMS allocations under the unconstrained setting. We also show that exact GMMS allocations are guaranteed to exist when the valuations of the agents are either binary or identical. We empirically show that GMMS allocations exist when the valuations are drawn from Gaussian and Uniform distributions. These results highlight that, for unconstrained settings, we do not fall short on generic existence results by strengthening the existing fairness notions. <br><br> <b>Application to Two-Sided Fair Recommendation Systems </b> <br>We investigate the problem of fair recommendation in two-sided online platforms, such as Amazon, Netflix, and Spotify, consisting of customers on one side and producers on the other. These services have typically focused only on maximizing customer satisfaction by tailoring the recommendations according to the preferences of individual customers, which may be detrimental for the producers. We consider fairness issues that span both customers and producers. Our approach involves a mapping of the fair recommendation problem to a constrained version of the fair allocation problem. Our proposed algorithm guarantees at least MMS exposure for most of the producers and EF1 fairness for every customer. We establish theoretical guarantees and provide empirical evidence through extensive evaluations on real-world datasets. <br><br> <b>Application to Classification Problems under Prior Probability Shifts </b> <br>We consider the problem of fair classification under prior probability shifts, which is a kind of distributional change occurring between the training and test datasets. Such shifts can be observed in the yearly records of several real-world datasets, such as COMPAS. If unaccounted for, such shifts can cause the predictions to become unfair towards specific population sub-groups. We define a fairness notion, called proportional equality (PE) which is motivated by solution concepts from the fair allocation literature, and accounts for prior probability shifts. We develop an algorithm CAPE that uses prevalence estimation techniques, sampling and an ensemble of classifiers to ensure fair predictions. We evaluate the performance of CAPE on real-world datasets and compare its performance with state-of-the-art fair algorithms. Our findings indicate that CAPE ensures PE-fair predictions, with low compromise on other performance metrics.<br/><br/> Link to the Online Colloquium:<br/> <a href="https://teams.microsoft.com/dl/launcher/launcher.html?url=%2f_%23%2fl%2fmeetup-join%2f19%3ameeting_NWFhN2ViMWQtY2M5MC00OTg0LWJiNjMtNDY3OTM0NjM2MmIx%40thread.v2%2f0%3fcontext%3d%257b%2522Tid%2522%253a%25226f15cd97-f6a7-41e3-b2c5-ad4193976476%2522%252c%2522Oid%2522%253a%2522d2ea60f8-0059-4875-aed5-2792decae1e5%2522%257d%26anon%3dtrue&type=meetup-join&deeplinkId=233760ea-e842-4141-a270-4e5d06a76da5&directDl=true&msLaunch=true&enableMobilePage=true&suppressPrompt=true">https://teams.microsoft.com/dl/launcher/launcher.html?url=%2f_%23%2fl%2fmeetup-join%2f19%3ameeting_NWFhN2ViMWQtY2M5MC00OTg0LWJiNjMtNDY3OTM0NjM2MmIx%40thread.v2%2f0%3fcontext%3d%257b%2522Tid%2522%253a%25226f15cd97-f6a7-41e3-b2c5-ad4193976476%2522%252c%2522Oid%2522%253a%2522d2ea60f8-0059-4875-aed5-2792decae1e5%2522%257d%26anon%3dtrue&type=meetup-join&deeplinkId=233760ea-e842-4141-a270-4e5d06a76da5&directDl=true&msLaunch=true&enableMobilePage=true&suppressPrompt=true</a> DTSTART:20200501T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200505T120000Z UID:b1f9e3f76533239cd3abe469b31a8cae-72 DTSTAMP:19700101T120009Z DESCRIPTION:Model Extraction defense using Modified Variational Autoencoder URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/72/model-extraction-defense-using-modified-variational-autoencoder/ SUMMARY:Machine Learning as a Service (MLaaS) exposes machine learning (ML) models that are trained on confidential datasets to users in the form of an Application Programming Interface (API). Since the MLaaS models are deployed for commercial purposes the API is available as a pay-per-query service. A malicious user or attacker can exploit these APIs to extract a close approximation of the MLaaS model by training a substitute model using only black-box query access to the API, in a process called model extraction. The attacker is restricted to extract the MLaaS model using a limited query budget because of the paid service. The model extraction attack is invoked by firing queries that belong to a substitute dataset that consists of either (i) Synthetic Non-Problem Domain (SNPD), (ii) Synthetic Problem Domain (SPD), or (iii) Natural Non-Problem Domain (NNPD) dataset. <br><br><br>In this work, we propose a novel defense framework against model extraction, using a hybrid anomaly detector composed of an encoder and a detector. In particular we propose a modified Variational Autoencoder, VarDefend, which uses a loss function, specially designed, to separate the encodings of queries fired by malicious users from those by benign users. We consider two scenarios: (i) stateful defense where an MLaaS provider stores the queries made by each client for discovering any malicious pattern, (ii) stateless defense where individual queries are discarded if they are flagged as out-of-distribution. Treating encoded queries from benign users as normal, one can use outlier detection models to identify encoded queries from malicious users in the stateless approach. For the stateful approach, a statistical test known as Maximum Mean Discrepancy (MMD) is used to match the distribution of the encodings of the malicious queries with those of the in-distribution encoded samples. In our experiments, we observed that our stateful defense mechanism can completely block one representative attack for each of the three types of substitute datasets, without raising a single false alarm against queries made by a benign user. The number of queries required to block an attack is much smaller than those required by the current state-of-the-art model extraction defense PRADA. Further, our proposed approach can block NNPD queries that cannot be blocked by PRADA. Our stateless defense mechanism is useful against a group of colluding attackers without significantly impacting benign users. Our experiments demonstrate that, for MNIST and FashionMNIST dataset, proposed stateless defense rejects more than 98% of the queries made by an attacker belonging to either SNPD, SPD or NNPD datasets while rejecting only about 0.05% of all the queries made by a benign user. Our experiments also demonstrate that the proposed stateless approach makes the MLaaS model significantly more robust to adversarial examples crafted using the substitute model by blocking transferability.<br><br> The defense will be conducted online. Please fill the following form before 4th May 5 PM to receive the Microsoft Teams meeting invite. <a href="https://forms.gle/Bw4Lj36uFuugZZFY7"> https://forms.gle/Bw4Lj36uFuugZZFY7 </a> DTSTART:20200505T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200504T120000Z UID:ff681728ca85ca1e5b1d243b64eb4a13-73 DTSTAMP:19700101T120010Z DESCRIPTION:Neural Graph Embedding Methods for Natural Language Processing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/73/neural-graph-embedding-methods-for-natural-language-processing/ SUMMARY:Graphs are all around us, ranging from citation and social networks to Knowledge Graphs (KGs). They are one of the most expressive data structures which have been used to model a variety of problems. Knowledge graphs are structured representations of facts in a graph, where nodes represent entities and edges represent relationships between them. Recent research has resulted in the development of several large KGs; examples include DBpedia, YAGO, NELL, and Freebase. However, all of them tend to be sparse with very few facts per entity. For instance, NELL KG consists of only 1.34 facts per entity. In the first part of the thesis, we propose three solutions to alleviate this problem: (1) KG Canonicalization, i.e., identifying and merging duplicate entities in a KG, (2) Relation Extraction which involves automating the process of extracting semantic relationships between entities from unstructured text, and (3) Link prediction which includes inferring missing facts based on the known facts in a KG. For KG Canonicalization, we propose CESI (Canonicalization using Embeddings and Side Information), a novel approach that performs canonicalization over learned embeddings of Open KGs. The method extends recent advances in KG embedding by incorporating relevant NP and relation phrase side information in a principled manner. For relation extraction, we propose RESIDE, a distantly-supervised neural relation extraction method which utilizes additional side information from KGs for improved relation extraction. Finally, for link prediction, we propose InteractE which extends ConvE, a convolutional neural network-based link prediction method, by increasing the number of feature interaction through three key ideas – feature permutation, a novel feature reshaping, and circular convolution. Through extensive experiments on multiple datasets, we demonstrate the effectiveness of our proposed methods.<br> <br> Traditional Neural Networks like Convolutional Networks and Recurrent Neural Networks are constrained to handle Euclidean data. However, graphs in Natural Language Processing (NLP) are prominent. Recently, Graph Convolutional Networks (GCNs) have been proposed to address this shortcoming and have been successfully applied for several problems. In the second part of the thesis, we utilize GCNs for Document Timestamping problem, which forms an essential component of tasks like document retrieval, and summarization. For this, we propose NeuralDater which leverages GCNs for jointly exploiting syntactic and temporal graph structures of document for obtaining state-of-the-art performance on the problem. We also propose SynGCN, a flexible Graph Convolution based method for learning word embeddings which utilize dependency context of a word instead of linear context for learning more meaningful word embeddings. In this third part of the thesis, we address two limitations of existing GCN models, i.e., (1) The standard neighborhood aggregation scheme puts no constraints on the number of nodes that can influence the representation of a target node. This leads to a noisy representation of hub-nodes which coves almost the entire graph in a few hops. To address this shortcoming, we propose ConfGCN (Confidence-based GCN) which estimates confidences to determine the importance of a node on another during aggregation, thus restricting its influence neighborhood. (2) Most of the existing GCN models are limited to handle undirected graphs. However, a more general and pervasive class of graphs are relational graphs where each edge has a label and direction associated with it. Existing approaches to handle such graphs suffer from over-parameterization and are restricted to the learning representation of nodes only. We propose CompGCN, a novel Graph Convolutional framework which jointly embeds entity and relations in a relational graph. CompGCN is parameter efficient and scales with the number of relations. It leverages a variety of entity-relation composition operations from KG Embedding techniques and achieves demonstrably superior results on node classification, link prediction, and graph classification tasks. DTSTART:20200504T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200424T120000Z UID:ea86454fafd0e48daafeadada7063867-74 DTSTAMP:19700101T120012Z DESCRIPTION:Algorithms for Social Good in Online Platforms with Guarantees on Honest Participation and Fairness URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/74/algorithms-for-social-good-in-online-platforms-with-guarantees-on-honest-participation-and-fairness/ SUMMARY:Recent decades have seen a revolution in the way people communicate, buy products, learn new things, and share life experiences. This has spurred the growth of online platforms that enable users from all over the globe to buy/review/recommend products and services, ask questions and provide responses, participate in online learning, etc.</br>There are certain crucial requirements that are required to be satisfied by the online forums for ensuring their trustworthiness and sustainability. In this thesis, we are concerned with three of these requirements: social welfare maximization, honest participation by the users, and fairness in decision making. In particular, we address three contemporary problems in online platforms and obtain principled solutions that achieve social welfare maximization while satisfying honest participation and fairness of allocation. The three problems considered are set in the context of three different platforms: online review or Q&A forums, online discussion forums, and online search platforms. In each case, we develop an abstraction of the problem and solve it in its generality.<br><b>Ballooning Multi-armed Bandits</b><br>In our first problem, we consider online platforms where the users are shown user generated content such as reviews on an e-commerce platform or answers on a Q&A platform. The number of reviews/answers increases over time. We seek to design an algorithm that quickly learns the best review/best answer and displays it prominently. We model this problem as a novel multi-armed bandit formulation (which we call ballooning bandits) in which the set of arms expands over time. We first show that when the number of arms grows linearly with time, one cannot achieve sub-linear regret. In a realistic special case, where the best answer is likely to arrive early enough, we prove that we can achieve optimal sublinear regret guarantee. We prove our results for best answer arrival time distributions that have sub-exponetal or sub-Pareto tails.<br><b>Strategy-proof Allocation of Indivisible Goods with Fairness Guarantees</b><br>Second, we consider the problem of fairness in online search platforms. We view the sponsored ad-slots on these platforms as indivisible goods to be allocated in a fair manner among competing advertisers. We use envy-freeness up to one good (EF1) and maximin fair share (MMS) allocation as the fairness notions. The problem is to maximize the overall social welfare subject to these fairness constraints. We first prove under a single parameter setting that the problem of social welfare maximization under EF1 is NP-hard. We complement this result by showing that any EF1 allocation satisfies an 1/2-approximation guarantee and present an algorithm with a (1, 1/2) bi-criteria approximation guarantee. We finally show in a strategic setting that one can design a truthful mechanism with the proposed fair allocation.<br><b>Coalition Resistant Credit Score Functions</b><br>In the third problem, we study manipulation in online discussion forums. We consider a specific but a common form of manipulation namely manipulation by coalition formation. We design a manipulation resistant credit scoring rule that assigns to each user a score such that forming a coalition is discouraged. In particular, we study the graph generated by the interactions on the platform and use community detection algorithms. We show that the community scores given by community detection algorithms that maximize modularity lead to a coalition resistant credit scoring rule. This in turn leads to sustainable discussion forums with honest participation from users, devoid of any coalitional manipulation. DTSTART:20200424T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200515T120000Z UID:a0df4facc81f83e36e1bc66b1bcf2320-75 DTSTAMP:19700101T120015Z DESCRIPTION:Decision Making under Uncertainty : Reinforcement Learning Algorithms and Applications in Cloud Computing, Crowdsourcing and Predictive Analytics URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/75/decision-making-under-uncertainty-reinforcement-learning-algorithms-and-applications-in-cloud-computing-crowdsourcing-and-predictive-analytics/ SUMMARY:In this thesis, we study both theoretical and practical aspects of decision making, with a focus on reinforcement learning based methods. Reinforcement learning (RL) is a form of semi-supervised learning in which the agent learns the decision making strategy by interacting with its environment. We develop novel reinforcement learning algorithms and study decision problems in the domains of cloud computing, crowdsourcing and predictive analytics.<br>In the first part of the thesis, we develop a model free reinforcement learning algorithm with faster convergence named Generalized Speedy Q-learning and analyze its finite time performance. This algorithm integrates ideas from the well-known Speedy Q-learning algorithm and the generalized Bellman equation to derive a simple and efficient update rule such that its finite time bound is better than that of Speedy Q-learning for MDPs with a special structure. Further, we extend our algorithm to deal with large state and action spaces by using function approximation.<br>Extending the idea in the above algorithm, we develop a novel Deep Reinforcement Learning algorithm by combining the technique of successive over-relaxation with Deep Q-networks. The new algorithm, named SOR-DQN, uses modified targets in the DQN framework with the aim of accelerating training. We study the application of SOR-DQN in the problem of auto-scaling resources for cloud applications, for which existing algorithms suffer from issues such as slow convergence, poor performance during the training phase and non-scalability.<br>Next, we consider an interesting research problem in the domain of crowdsourcing - that of efficiently allocating a fixed budget among a set of tasks with varying difficulty levels. Further, the assignment of tasks to workers with different skill levels is tackled. This problem is modeled in the RL framework and an approximate solution is proposed to deal with the exploding state space.We also study the following problem in predictive analytics : predicting the future values of system parameters well in advance for a large-scale software or industrial system, which is important for avoiding disruptions. An equally challenging and useful exercise is to identify the "important" parameters and optimize them in order to attain good system performance. In addition to devising an end-to-end solution for the problem, we present a case study on a large-scale enterprise system to validate the effectiveness of the proposed approach. DTSTART:20200515T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200515T120000Z UID:edb3028b88a7e06729148e4fc8ca964f-78 DTSTAMP:19700101T120009Z DESCRIPTION:Online Learning from Relative Subsetwise Preferences URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/78/online-learning-from-relative-subsetwise-preferences/ SUMMARY:The elicitation and aggregation of preferences is often the key to making better decisions. Be it a perfume company wanting to relaunch their 5 most popular fragrances, a movie recommender system trying to rank the most favoured movies, or a pharmaceutical company testing the relative efficacies of a set of drugs, learning from preference feedback is a widely applicable problem to solve. One can model the sequential version of this problem using the classical multiarmed-bandit (MAB) (e.g., Auer, 2002) by representing each decision choice as one bandit-arm, or more appropriately as a Dueling-Bandit (DB) problem (Yue and Joachims, 2009). Although DB is similar to MAB in that it is an online decision making framework, DB is different in that it specifically models learning from pairwise preferences. In practice, it is often much easier to elicit information, especially when humans are in the loop, through relative preferences: Item A is better than item B is easier to elicit than its absolute counterpart: Item A is worth 7 and B is worth 4. <br><br> However, instead of pairwise preferences, a more general subset-wise preference model is more relevant in various practical scenarios, e.g. recommender systems, search engines, crowd-sourcing, e-learning platforms, design of surveys, ranking in multiplayer games. Subset-wise preference elicitation is not only more budget friendly, but also flexible in conveying several types of feedback. For example, with subset-wise preferences, the learner could elicit the best item, a partial preference of the top 5 items, or even an entire rank ordering of a subset of items, whereas all these boil down to the same feedback over pairs (subsets of size 2). The problem of how to learn adaptively with subset-wise preferences, however, remains largely unexplored; this is primarily due to the computational burden of maintaining a combinatorially large, O(n^k), size of preference information in general. <br><br> We take a step in the above direction by proposing Battling Bandits (BB)---a new online learning framework to learn a set of optimal (good) items by sequentially, and adaptively, querying subsets of items of size up to k (k>=2). The preference feedback from a subset is assumed to arise from an underlying parametric discrete choice model, such as the well-known Plackett-Luce model, or more generally any random utility (RUM) based model. It is this structure that we leverage to design efficient algorithms for various problems of interest, e.g. identifying the best item, set of top-k items, full ranking etc., for both in PAC and regret minimization setting. We propose computationally efficient and (near-) optimal algorithms for above objectives along with matching lower bound guarantees. Interestingly this leads us to finding answers to some basic questions about the value of subset-wise preferences: Does playing a general k-set really help in faster information aggregation, i.e. is there a tradeoff between subsetsize-k vs the learning rate? Under what type of feedback models? How do the performance limits (performance lower bounds) vary over different combinations of feedback and choice models? And above all, what more can we achieve through BB where DB fails? <br><br> We proceed to analyse the BB problem in the contextual scenario – this is relevant in settings where items have known attributes, and allows for potentially infinite decision spaces. This is more general and of practical interest than the finite-arm case, but, naturally, on the other hand more challenging. Moreover, none of the existing online learning algorithms extend straightforwardly to the continuous case, even for the most simple Dueling Bandit setup (i.e. when k=2). Towards this, we formulate the problem of Contextual Battling Bandits (C-BB) under utility based subsetwise-preference feedback, and design provably optimal algorithms for the regret minimization problem. Our regret bounds are also accompanied by matching lower bound guarantees showing optimality of our proposed methods. All our theoretical guarantees are corroborated with empirical evaluations. <br><br> Lastly, it goes without saying, that there are still many open threads to explore based on BB. These include studying different choice-feedback model combinations, performance objectives, or even extending BB to other useful frameworks like assortment selection, revenue maximization, budget-constrained bandits etc. Towards the end we will also discuss some interesting combinations of the BB framework with other, well-known, problems, e.g. Sleeping / Rotting Bandits, Preference based Reinforcement Learning, Learning on Graphs, Preferential Bandit-Convex-Optimization etc. <br><br> Microsoft Teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzFmNTdmODYtYjhhZi00Yjc4LTg3NWItNmEyNzc5NzlkMzQ1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22620fb6db-36c5-4f95-ba45-6ed8e824aa28%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzFmNTdmODYtYjhhZi00Yjc4LTg3NWItNmEyNzc5NzlkMzQ1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22620fb6db-36c5-4f95-ba45-6ed8e824aa28%22%7d</a> DTSTART:20200515T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200528T120000Z UID:c5df934edc8818c128e58e067989ea4a-81 DTSTAMP:19700101T120014Z DESCRIPTION:Fault Aware Read-Copy-Update URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/81/fault-aware-read-copy-update/ SUMMARY:Deferred freeing is the fundamental technique used in Read-Copy-Update (RCU) synchronization technique where reclamation of resources is deferred until the completion of all active RCU read-side critical sections. We observe that faults inside an RCU read-side critical section can indefinitely block writers that are waiting for the completion of RCU readers and also lead to system failures by preventing the reclamation of deferred resources. We show that the impact of such faults in the Linux kernel is global; a fault in one subsystem can propagate and exhaust critical resources in other unrelated subsystems opening a window of opportunity for DoS-based attacks. For example, a fault in a filesystem can exhaust the process ulimit resulting in fork failures. Since, guaranteeing the absence of faults is practically impossible, it is imperative to harden RCU to tolerate faults. We first study the impact of mitigating lockup by termination of the faulty thread, as thread termination is standard approach used by Linux as recovery strategy. Whereas, another solution is stack based and do not require termination of faulty thread. We demonstrate the impact of faults in RCU read-side critical sections and present RCU recovery techniques that use novel approaches to detect and isolate effect of such faults. We also discuss system consistency once the fault is handled by our approaches. Timely recovery results in a usable system, preserving the user application state and increasing the system’s availability. Our evaluation in the Linux kernel shows that our solution can prevent resource exhaustion in the presence of faults with no additional overhead in the absence of faults. <br> <br> Teams Meeting Link:<br> <a href="https://teams.microsoft.com/dl/launcher/launcher.html?url=%2f_%23%2fl%2fmeetup-join%2f19%3ameeting_Nzk2YTg2ZDEtOWRhYy00OGYyLThmZTYtZWI5NTZiMGY0YzRl%40thread.v2%2f0%3fcontext%3d%257b%2522Tid%2522%253a%25226f15cd97-f6a7-41e3-b2c5-ad4193976476%2522%252c%2522Oid%2522%253a%252247d9ed45-e131-49b4-9b89-ac82d3c3da13%2522%257d%26anon%3dtrue&type=meetup-join&deeplinkId=d862ca07-3513-4597-8b62-e5bea2346d12&directDl=true&msLaunch=true&enableMobilePage=true&suppressPrompt=true">https://teams.microsoft.com/dl/launcher/launcher.html?url=%2f_%23%2fl%2fmeetup-join%2f19%3ameeting_Nzk2YTg2ZDEtOWRhYy00OGYyLThmZTYtZWI5NTZiMGY0YzRl%40thread.v2%2f0%3fcontext%3d%257b%2522Tid%2522%253a%25226f15cd97-f6a7-41e3-b2c5-ad4193976476%2522%252c%2522Oid%2522%253a%252247d9ed45-e131-49b4-9b89-ac82d3c3da13%2522%257d%26anon%3dtrue&type=meetup-join&deeplinkId=d862ca07-3513-4597-8b62-e5bea2346d12&directDl=true&msLaunch=true&enableMobilePage=true&suppressPrompt=true</a> DTSTART:20200528T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200529T120000Z UID:df269f512cde01a7f8dfdb2000fdeb0d-82 DTSTAMP:19700101T120014Z DESCRIPTION:Representing Networks: Centrality, Node Embeddings, Community Outliers and Graph Representation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/82/representing-networks-centrality-node-embeddings-community-outliers-and-graph-representation/ SUMMARY:We start our technical work in this thesis by exploring the classical concept of node centrality (also known as influence measure) in information network. Like clustering, node centrality is also an ill-posed problem. There exist several heuristics and algorithms to compute the centrality of a node in a graph, but there is no formal definition of centrality available in the network science literature. Lately, researchers have proposed axiomatic frameworks for the centrality of a node in a network. However, these existing formal frameworks are not generic in nature in terms of characterizing the space of influence measures in complex networks. In this work, we propose a set of six axioms in order to capture most of the intrinsic properties that any influence measure ideally should satisfy. We also characterize existing measures of centrality with respect to this framework. <br> Next, we focus more on the representation learning on networks. Network embedding is required as real life networks are large, extremely sparse and discrete in nature. We investigate the problem of unsupervised node representation in attributed networks through informative random walk. Edges are also useful for various downstream network mining tasks, but most of the existing homogeneous network representation learning approaches focus on embedding the nodes of a graph. So, we propose a novel unsupervised algorithm to embed the edges of a network, through the application of the classical concept of line graph of a network. The optimization framework of edge embedding connects to the concept of node centrality in the representation learning framework. Finally, we also conduct research on attributed hypergraphs. We propose a novel graph neural network to represent and classify hypernodes. <br> Outlier analysis (or anomaly detection) is another important problem for the network science community. All the real-world networks contain outlier nodes to some extent. Empirically we have shown that outliers can affect the quality of network embedding if not handled properly. So, we integrate the process of network embedding and outlier detection into a single framework. In this research thread, we first propose a matrix factorization based approach which minimizes the effect of outlier nodes in the framework of attributed network embedding. Next, we propose two neural network architectures, based on L2 regularization and adversarial training respectively, to minimize the effect of outliers on node embedding of an attributed network. Further, extending the concept of support vector data description, we propose a novel algorithm which integrates node embedding, community detection and outlier detection into a single optimization framework by exploiting the link structure of a graph. <br> So far, we have conducted research only on the individual components of a graph, i.e., on nodes and edges. In the last part of the thesis, we focus on graph level representation and tasks. First, we propose a supervised graph neural network based algorithm with hierarchical pooling strategy to classify a graph from a set of graphs. Next, we propose a novel GNN based algorithm for the unsupervised representation of a graph from a set of graphs, so that similar graphs are represented closely in the embedding space and dissimilar graphs are separated away. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/team/19%3ad5d87af5b08d4f14b81c06c903932960%40thread.tacv2/conversations?groupId=ef37e047-58df-4367-b37e-9bb717bb42bc&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/team/19%3ad5d87af5b08d4f14b81c06c903932960%40thread.tacv2/conversations?groupId=ef37e047-58df-4367-b37e-9bb717bb42bc&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20200529T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200601T120000Z UID:e5a87c555b7d49f090eef7db551aaebc-83 DTSTAMP:19700101T120011Z DESCRIPTION:Learning-Based Controlled Concurrency Testing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/83/learning-based-controlled-concurrency-testing/ SUMMARY:Concurrency bugs are notoriously hard to detect and reproduce. Controlled concurrency testing (CCT) techniques aim to offer a solution, where a scheduler explores the space of possible interleavings of a concurrent program looking for bugs. Since the set of possible interleavings is typically very large, these schedulers employ heuristics that prioritize the search to "interesting" subspaces. However, current heuristics are typically tuned to specific bug patterns, which limits their effectiveness in practice. <br><br> In this work, we present QL, a learning-based CCT framework where the likelihood of an action being selected by the scheduler is influenced by earlier explorations. We leverage the classical Q-learning algorithm to explore the space of possible interleavings, allowing the exploration to adapt to the program under test, unlike previous techniques. We have implemented and evaluated QL on a set of microbenchmarks, complex protocols, as well as production cloud services. In our experiments, we found QL to consistently outperform the state-of-the-art in CCT. <br><br> This is joint work with Pantazis Deligiannis (Microsoft Research), Arpita Biswas (Indian Institute of Science) and Akash Lal (Microsoft Research). <br><br> Microsoft Teams Meeting Link: <br> <br><a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjIwMzU5NDItM2Q2ZC00Zjg5LTkzYTYtMDVkODg2M2I0OGYw%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224e94f9c8-085e-46c8-b31f-468b334d3138%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjIwMzU5NDItM2Q2ZC00Zjg5LTkzYTYtMDVkODg2M2I0OGYw%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224e94f9c8-085e-46c8-b31f-468b334d3138%22%7d</a> DTSTART:20200601T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200529T120000Z UID:1f41c77d0edaf9acdbfba3b344742f58-84 DTSTAMP:19700101T120015Z DESCRIPTION:Constant-rate Non-malleable Codes and their Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/84/constant-rate-non-malleable-codes-and-their-applications/ SUMMARY:Non-malleable codes (NMCs) introduced by Dziembowski, Pietrzak and Wichs in ITCS 2010, provide powerful security guarantees where standard error-correcting codes can not provide any guarantee: a decoded message is either the same or completely independent of the underlying message. NMCs have found applications to various aspects of cryptography such as CCA secure encryption, tamper and leakage resilient cryptography, non-malleable commitments, non-malleable secret sharing schemes and so on. <br><br> In this talk, we present an application of NMCs to the fascinating problem of NMCs to the fascinating problem of Privacy Amplification. In the problem of privacy amplification, two parties, Alice and Bob, who a-priori share a weak secret, to agree on a uniform secret key, in the presence of a computationally unbounded adversary Eve. Building privacy amplification protocols with constant entropy loss and constant round complexity was open since 1988 (and recently closed in an independent work of Li [CCC 19]). In this talk, we will show how to construct such a privacy amplification protocol under the existence of non-malleable code with certain strong security guarantees. Next, we will also discuss the first explicit construction of a constant rate, constant state non-malleable code. <br><br> This talk is based on joint works with Eshan Chattopadhyay, Bhavana Kanukurthi and Sruthi Sekar. <br><br> References: [1] Bhavana Kanukurthi, Sai Lakshmi Bhavana Obbattu, and Sruthi Sekar. Four-state non-malleable codes with explicit constant rate. In Theory of Cryptography Conference, TCC 2017. Invited to Journal of Cryptology. <br> [2] Eshan Chattopadhyay, Bhavana Kanukurthi, Sai Lakshmi Bhavana Obbattu, and Sruthi Sekar. Privacy amplification from non-malleable codes. In Indocrypt 2019. <br> <br> <br> Please fill this form:<a href="https://docs.google.com/forms/d/e/1FAIpQLSdYuJUadeb9EU60_G_-dm8iy7Y0GjBqzaE8JWr-UL8G9KgTqA/viewform?usp=sf_link"> https://docs.google.com/forms/d/e/1FAIpQLSdYuJUadeb9EU60_G_-dm8iy7Y0GjBqzaE8JWr-UL8G9KgTqA/viewform?usp=sf_link</a> by 28 May to be added to the conversation DTSTART:20200529T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200615T120000Z UID:46ad8380ed44fdfc1ba21edb21424e97-85 DTSTAMP:19700101T120011Z DESCRIPTION:Privacy Preserving Machine Learning via Multi-party Computation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/85/privacy-preserving-machine-learning-via-multi-party-computation/ SUMMARY:Privacy-preserving machine learning (PPML) via Secure Multi-party Computation (MPC) has gained momen-tum in the recent past. Assuming a minimal network of pair-wise private channels, we propose an efficient four-party PPML framework over rings, FLASH, the first of its kind in the regime of PPML framework, that achieves the strongest security notion of Guaranteed Output Delivery (all parties obtain the output irrespective of adversary's behaviour). The state of the art ML frameworks such as ABY3 by Mohassel et.al (ACM CCS'18) and SecureNN by Wagh et.al (PETS'19) operate in the setting of 3 parties with one malicious corruption but achieve the weaker security guarantee of abort. We demonstrate PPML with real-time efficiency, using the following custom-made tools that overcome the limitations of the aforementioned state-of-the-art-- (a) dot prod-uct, which is independent of the vector size unlike the state-of-the-art ABY3, SecureNN and ASTRA by Chaudhari et.al (ACM CCSW'19), all of which have linear dependence on the vector size. (b) Truncation which is constant round and free of circuits like Ripple Carry Adder (RCA), unlike ABY3 which uses these circuits and has round complexity of the order of depth of these circuits. We then exhibit the application of our FLASH framework in the secure server-aided prediction of vital algorithms: Linear Regression, Logistic Regression, Deep Neural Networks, and Binarized Neural Networks. We substantiate our theoretical claims through im-provement in benchmarks of the aforementioned algorithms when compared with the current best framework ABY3. All the protocols are implemented over a 64-bit ring in LAN and WAN. Our experiments demonstrate that, for MNIST dataset, the improvement (in terms of throughput) ranges from 11x to 1390x over Local Area Network (LAN) and Wide Area Network (WAN) together. <br> <br> Online link to join Microsoft meeting: <br> <a href="https://teams.microsoft.com/_#/pre-join-calling/19:meeting_MmMxMDZkNmItZWZkOS00ZGJhLTgyYzYtNjlhODZiYjk5NzNj@thread.v2">https://teams.microsoft.com/_#/pre-join-calling/19:meeting_MmMxMDZkNmItZWZkOS00ZGJhLTgyYzYtNjlhODZiYjk5NzNj@thread.v2</a> DTSTART:20200615T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200624T120000Z UID:1e8dbccd518ba22c184306fbbf33bdd0-86 DTSTAMP:19700101T120013Z DESCRIPTION:Embedding Networks: Node and Graph Level Representations URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/86/embedding-networks-node-and-graph-level-representations/ SUMMARY:Network representation learning is important to carry out various network analysis downstream tasks. Graphs are the most suitable structures to represent relational data such as social networks and molecular structures. In this thesis work, we focus on learning representations of the nodes as well as of the entire graphs. Graph neural networks got significant importance for graph representation. Recently, attention mechanisms on graphs show promising results for classification tasks. Most of the attention mechanisms developed in graph literature use attention to derive the importance of a node or a pair of nodes for different tasks. But in the real world situation, calculating importance up to a pair of nodes is not adequate. <br> To address this problem, we introduce a novel GNN based approach, subgraph attention, to classify the nodes of a graph. On the other hand, the hierarchical graph pooling is promising in the recent literature. But, not all the hierarchies of a graph play an equal role for graph classification. Towards this end, we propose an algorithm called SubGattPool to find the important nodes in a hierarchy and the importance of individual hierarchies in a graph for embedding and classifying the graphs given a collection of graphs. Moreover, existing pooling approaches do not consider both the region based as well as the graph level importance of the nodes together. In the next research work, we solve this issue by proposing a novel pooling layer named R2pool which retains the most informative nodes for the next coarser version of the graph. Further, we integrate R2pool with our branch training strategy to learn coarse to fine representations and improve the model's capability for graph classification by exploiting multi-level prediction strategy. Thorough experimentation on both the real world and synthetic graphs shows the merit of the proposed algorithms over the state-of-the-art. <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/team/19%3a35aa9398569340ed9061a35f0589ffe2%40thread.tacv2/conversations?groupId=d9e50dc5-3600-46c1-8888-998889fcedb8&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/team/19%3a35aa9398569340ed9061a35f0589ffe2%40thread.tacv2/conversations?groupId=d9e50dc5-3600-46c1-8888-998889fcedb8&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20200624T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200717T120000Z UID:e81438946269fc667d8cdbdbb867c327-87 DTSTAMP:19700101T120011Z DESCRIPTION:Modern Combinatorial Optimization Problems: Balanced Clustering and Fair Knapsack URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/87/modern-combinatorial-optimization-problems-balanced-clustering-and-fair-knapsack/ SUMMARY:In many classical optimization problems, it is desirable to have an output that is equitable in some sense. The property equitability could be defined differently for different optimization problems. We study this property in two classical optimization problems, clustering and knapsack. In the clustering problem, we desire to have a cost of the clustering evenly distributed among the clusters. We study this problem under the name cost-balanced k-clustering. In the knapsack problem, we desire to have a packing which is fair in some sense. We study this problem under the name fair knapsack. In most of the clustering objectives like k-median or k-center, the cost of assigning a client to the cluster is considered to be borne by client. Algorithms optimizing such objectives might output a solution where few clusters have very large cost and few clusters have very small cost. Cost-balanced k-clustering problem aims to obtain the clustering which is cost balanced. We consider the objective of minimizing the maximum cost of each cluster, where the cost of a cluster is the sum of distances of all the points in that cluster from the center of that cluster. We define the notion of γ-stability, for γ > 1, for the problem and give a poly time algorithm for 1.5-stable instances of the problem. The algorithm requires an optimal value as an input. We also modify this algorithm for 1.5+eps, eps>0, stable instance that does not require an optimal value as an input. We also define the more general version of the problem and name it "lp" cost-balanced k-clustering. Given a black-box algorithm which gives constant factor approximation to the "lp" cost k-clustering problem, we show a procedure that runs in poly time and gives bi-criteria approximation to the "lp" cost-balanced k-clustering problem.<br> In this work, we also consider the notion of group fairness in the knapsack problems. In this setting, each item belongs to some category, and our goal is to compute a subset of items such that each category is fairly represented, in addition to the total weight of the subset not exceeding the capacity, and the total value of the subset being maximized. We study various notions of group fairness, such as the number of items from each category, the total value of items from each category, and the total weight of items from each category. We give algorithms and hardness results for these problems. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/team/19%3abcdacd2f9b9b4f4e8534f288b0598817%40thread.tacv2/conversations?groupId=d79f4d99-3e9e-4035-a301-09c28012be9e&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/team/19%3abcdacd2f9b9b4f4e8534f288b0598817%40thread.tacv2/conversations?groupId=d79f4d99-3e9e-4035-a301-09c28012be9e&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20200717T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200721T120000Z UID:85cd50ba2df5949bc733c05586357005-88 DTSTAMP:19700101T120014Z DESCRIPTION:Equivalence test for the trace iterated matrix multiplication polynomial URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/88/equivalence-test-for-the-trace-iterated-matrix-multiplication-polynomial/ SUMMARY:An m-variate polynomial f is affine equivalent to an n-variate polynomial g if m > n and there is a full rank n * m matrix A and a n-dimensional vector b such that f(x) = g(Ax + b). Given blackbox access to f and g (i.e membership query access) the affine equivalence test problem is to determine whether f is affine equivalent to g, and if yes then output a rank n matrix A and a vector b such that f(x) = g(Ax + b). This problem is at least as hard as graph isomorphism and algebra isomorphism even when the coefficients of f and g are given explicitly (Agarwal and Saxena, STACS 2006), and has been studied in literature by fixing g to be some interesting family of polynomials. In this work, we fix g to be the trace of the product of d, w * w symbolic matrices. We call this polynomial 'Trace Iterated Matrix Multiplication' polynomial (Tr-IMM). Kayal, Nair, Saha and Tavenas (CCC 2017) gave an efficient (i.e polynomial in m,w,d) randomized algorithm for the affine equivalence test of the 'Iterated Matrix Multiplication' polynomial (IMM), which is the (1,1)-th entry of the product of these symbolic matrices. Although the definitions of IMM and Tr-IMM are closely related and their circuit complexities are very similar, it is not clear whether an efficient affine equivalence test algorithm for IMM implies the same for Tr-IMM. In this thesis, we take a step towards showing that equivalence test for IMM and Tr-IMM have different complexity. We show that equivalence test for Tr-IMM reduces in randomized polynomial time to equivalence test for the determinant polynomial (Det), under mild conditions on the underlying field. If the converse is also true then equivalence tests for Tr-IMM and Det are randomized polynomial time equivalent. It would then follow from the work of Gupta, Garg, Kayal and Saha (ICALP 2019) that equivalence test for Tr-IMM over rationals is at least as hard as Integer Factoring. This would then be in sharp contrast with the complexity of equivalence test for Tr-IMM over rationals which can be solved efficiently in randomized polynomial time (by Kayal, Nair, Saha and Tavenas (CCC 2017)). <br> Recent Update: Soon after the thesis is written, we (together with Vineet Nair) have succeeded in showing the converse direction. So, the above conclusion is indeed true! This work has been accepted at the 45th International Symposium on Mathematical Foundations of Computer Science (MFCS 2020). <br> <br> Please fill out this form by July 19, 2020 to receive the Google Meet link: <br> <a href="https://docs.google.com/forms/d/e/1FAIpQLSfWTKc-z5v_dpOpfzb8JwZKFJVl5UN6aUR5Pw2yIwK22QCqww/viewform?usp=sf_link">https://docs.google.com/forms/d/e/1FAIpQLSfWTKc-z5v_dpOpfzb8JwZKFJVl5UN6aUR5Pw2yIwK22QCqww/viewform?usp=sf_link</a> DTSTART:20200721T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200722T120000Z UID:9a6f8c0fa90be70651fa0d5ea4e4475a-89 DTSTAMP:19700101T120012Z DESCRIPTION:Hypergraph Network Models: Learning, Prediction, and Representation in the Presence of Higher-Order Relations URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/89/hypergraph-network-models-learning-prediction-and-representation-in-the-presence-of-higher-order-relations/ SUMMARY:The very thought about relating objects makes us assume the relation would be pairwise, and not of a higher-order -- involving possibly more than two of them at a time. Yet in reality, higher-order relations do exist and are spread across multiple domains: medical science (e.g., co-existing diseases/symptoms), pharmacology (e.g., reacting chemicals), bibliometrics (e.g., collaborating researchers), the film industry (e.g., cast/crew), human resource (e.g., a team), social sciences (e.g., negotiating/conflicting nations), and so on. Since a collection of intersecting higher-order relations lose context when represented by a graph, "hypergraphs" -- graph-like structures that allow edges (called hyperedges/hyperlinks) spanning possibly more than two nodes -- capture them better. In a quest to better understand such relations, in this thesis we focus on solving a few network-science oriented problems involving hypergraphs. <br> Hypergraphs and Pairwise Links: In the first of three broad parts, we study the behavior of usual graph-oriented networks that have an otherwise-ignored hypergraph underpinning. We particularly establish the skewness a hypergraph introduces into its induced graphs, and the effect of these biases on the structure and evaluation of the well-known problem of link prediction in networks. We find that an underlying hypergraph structure makes popular heuristics such as common-neighbors overestimate their ability to predict links. Gathering enough evidence -- both theoretical and empirical -- to support the need to reestablish the evaluations of link prediction algorithms on hypergraph-derived networks, we propose adjustments that essentially undo the undesired effects of hypergraphs in performance scores. Motivated by this observation, we extend graph-based structural node similarity measures to cater to hypergraphs (although still, for similarity between pairs of nodes). To be specific, we first establish mathematical transformations that could transfer any graph-structure-based notion of similarity between node pairs to a hypergraph-structure-based one. Using exhaustive combinations of the newly established scores with the existing ones, we could show improvements in the performance of both structural as well as temporal link prediction. <br> "Predicting Higher-order Relations": For the second part of our thesis, we turn our attention towards a more central problem in hypergraphs -- the hyperlink/hyperedge prediction problem. It simply refers to developing models to predict the occurrence of missing or future hyperedges. We first study the effect of negative sampling (sampling the negative class) -- an exercise performed due to the extreme intractability of the set of all non-hyperlinks, also known as the class imbalance problem -- on hyperlink prediction, which has never been studied in the past. Since we observe hyperlink prediction algorithms performing differently under different negative sampling techniques, our experiments help the seemingly unimportant procedure gain some significance. Moreover, we contribute towards two benchmark negative sampling algorithms that would help standardize the step. Moving on from the negative sampling problem to predicting hyperlinks themselves, we work on two different approaches: a clique-closure based approach, and a sub-higher-order oriented one. While in the former, we develop and successfully test the clique-closure hypothesis -- that hyperlinks mostly form from cliques or near-cliques -- and are able to utilize it for hyperlink prediction via matrix completion (C3MM), the latter approach works on a novel information flow model in hypergraphs. More precisely, we introduce the concept of sub-hyperedges to capture the sub-higher-order notion in relations, and utilize an attention-based neural network model called SHONeN focusing on sub-hyperedges of a hyperedge. Owing to SHONeNs computational complexity, we propose a sub-optimal heuristic that is able to perform better than its baselines on the downstream task of predicting hyperedges. <br> "Higher-order Bipartite Relations": The third and final part of our thesis is dedicated exclusively to bipartite hypergraphs: structures that are used to capture higher-order relations between two disjoint node sets, e.g., a patients diagnosis (possibly multiple diseases and symptoms), a movie project (multiple actors and crew members), etc. We first capture the structure of real-world such networks using per-fixed bipartite hypergraphs (those where the set of left and right hyperedges is fixed beforehand), and then focus on the bipartite hyperlink prediction problem. Since existing self-attention based approaches meant for usual hypergraphs do not work for bipartite hypergraphs -- a fact that our experiments validate, we propose a cross-attention model for the same, and use the notion of set-matching over collections of sets to solve for bipartite hyperlink prediction. As a result, we develop a neural network architecture called CATSETMAT that performs way better than any of the other approaches meant to solve the bipartite hyperlink prediction problem. Last but not least, we also explain how, owing to an observation we call the positive-negative dilemma, existing state-of-the-art algorithms fail on bipartite hypergraphs. <br> <br> Microsoft Teams Meeting Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWQyMTY2OTMtMmI4MC00ZThiLThiN2QtNGU1M2FlYTExZDRi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227cc8bd44-135e-4281-b033-80d7f9df42fd%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWQyMTY2OTMtMmI4MC00ZThiLThiN2QtNGU1M2FlYTExZDRi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227cc8bd44-135e-4281-b033-80d7f9df42fd%22%7d</a> DTSTART:20200722T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200820T120000Z UID:85ea0464cbe830b966e4db6656b7b47a-90 DTSTAMP:19700101T120014Z DESCRIPTION:Verification of a Generative Separation Kernel URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/90/verification-of-a-generative-separation-kernel/ SUMMARY:In this thesis we present a technique to formally verify the functional correctness of template-based generative systems. A generative system is one in which a specialised system is generated for each given input specification. In a template-based generative system, the generated systems have a common template code, and the components of this template are generated based on the given input specification. Many safety-critical systems today are of this nature, with an important class being that of separation kernels. A separation kernel is a small specialized microkernel that provides a sand-boxed execution environment for a given set of processes called "subjects", and are commonly used in aerospace and defence applications. <br> The verification goal in the context of generative systems is to show that for every valid input specification the generated system refines an abstract specification that is generated according to the input specification. <br> The key stepping stone in our theory is the notion of a parametric Abstract Data Type or "machine". These are similar to classical machines except that they have parameters on which the behaviour of the machine depends. We propose a notion of refinement for such machines, which we call conditional parametric refinement, which essentially says that whenever the parameter values of the abstract and concrete parametric machines satisfy a given condition, the resulting machines enjoy a refinement relation. We define refinement conditions for this notion of refinement and show that they are sound and complete for total and deterministic machines. The refinement conditions are similar in structure to the classical case and can be discharged using standard Floyd-Hoare logic in a programmatic setting. <br> We use this framework to propose a two-step verification approach for template-based generative systems. Such a system naturally corresponds to a parametric machine, where the template corresponds to the definition of operations of the machine and generated components correspond to values for the parameters of the machine. In a similar way we construct an abstract parametric machine whose parameter values are generated based on the input specification. The first step of our approach is independent of the input specification and checks that the that the concrete parametric machine conditionally refines the abstract parametric machine, subject to a condition C on the parameter values generated. In the second step, which is input-specific, we check that for a given input specification, the generated abstract and concrete parameter values actually satisfy the condition C. Whenever this check passes for a given input specification, we can say that the generated system refines the abstract system. This gives us an effective verification technique for verifying generative systems that lies somewhere between verifying the generator and translation validation. <br> We demonstrate the effectiveness of our technique by applying it to verify the Muen Separation Kernel. Muen is an open-source template-based generative separation kernel which uses hardware support for virtualization to implement separation of subjects. We chose to model the virtualization layer (in this case Intel's VT-x layer) along with the rest of the hardware components like registers and memory, programmatically in software. Using the templates which Muen generator utilizes we constructed a parametric machine and carried out the first step of conditional parametric refinement for Muen, using the Spark Ada verification tool. This was a substantial effort involving about 20K lines of source code and annotation. We have also implemented a tool that automatically and efficiently performs the Step 2 check for a given separation kernel configuration. The tool is effective in proving the assumptions, leading to machine-checked proofs of correctness for 16 different input configurations, as well as in detecting issues like undeclared sharing of memory components in some seeded faulty configurations. <br> <br> <br> Details for joining online on Google Meet: <br> Link: <a href="https://meet.google.com/xkj-bugw-cxf">https://meet.google.com/xkj-bugw-cxf</a> <br> Please note that the meeting will be recorded as per Institute requirements. By joining the link you are giving your consent to the recording. DTSTART:20200820T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200828T120000Z UID:98aa8f9cf7b6e951295282260f19c437-91 DTSTAMP:19700101T120009Z DESCRIPTION:Scalable and Effective Polyhedral Auto-transformation Without using Integer Linear Programming URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/91/scalable-and-effective-polyhedral-auto-transformation-without-using-integer-linear-programming/ SUMMARY:In recent years, polyhedral auto-transformation frameworks have gained significant interest in general-purpose compilation, because of their ability to find and compose complex loop transformations that extract high performance from modern architectures. These frameworks automatically find loop transformations that either enhance locality, parallelism and minimize latency or a combination of these. Recently, focus has also shifting on developing intermediate representations, like MLIR, where complex loop transformations and data-layout optimizations can be incorporated efficiently in a single common infrastructure. <br> Polyhedral auto-transformation frameworks typically rely on complex Integer Linear Programming (ILP) formulations to find affine loop transformations. However, construction and solving these ILP problems is time consuming which increases compilation time significantly. Secondly, loop fusion heuristics in these auto-transformation frameworks are ad hoc, and modeling loop fusion efficiently would further degrade compilation time. <br> In this thesis, we first relax the ILP formulation in the Pluto algorithm. We show that even though LP relaxation reduces the time complexity of the problem, it does not reduce the compilation time significantly because of the complex construction of constraints. We also observe that due to relaxation, sub-optimal loop transformations that result in significant performance degradation may be obtained. Hence, we propose a new polyhedral auto-transformation framework, called Pluto-lp-dfp, that finds efficient affine loop transformations quickly, while relying on Plutos cost function. The framework decouples auto-transformation into three components: (1) loop fusion and permutation (2) loop scaling and shifting and (3) loop skewing components. In each phase, we solve a Linear Programming (LP) formulation instead of an ILP, thereby resulting in a polynomial time affine transformation algorithm. We propose a data structure, called fusion conflict graph, that allows us to model loop fusion to work in tandem with loop permutations, loop scaling and loop shifting transformations. We describe three greedy fusion heuristics, namely, max-fuse, typed-fuse and hybrid-fuse, of which, the hybrid-fuse and typed-fuse models incorporate parallelism preserving fusion heuristic without significant compilation time overhead. We also provide a characterization of time-iterated stencils that have tile-wise concurrent start and employ a different fusion heuristic in such programs. In our experiments, we demonstrate that Pluto-lp-dfp framework not only finds loop transformations quickly, resulting in significant improvements in compilation time, but also outperforms state-of-the-art polyhedral auto-parallelizers in terms of execution time of the transformed program. We observe that Pluto-lp-dfp is faster than PoCC and Pluto by a geomean factor of 461x and 2.2x in terms of compilation time. On larger NAS benchmarks, Pluto-lp-dfp was faster than Pluto by 246x. PoCC failed to find a transformation in a reasonable amount of time in these cases. In terms of execution time, the hybrid-fuse variant in Pluto-lp-dfp outperforms PoCC by a geomean factor 1.8x, with over 3x improvements in some cases. We also observe that Pluto-lp-dfp is faster than an improved version of Pluto by a factor of 7%, with a maximum performance improvement of 2x. <br> <br> Link to join online on Google Meet: <a href="https://meet.google.com/ioe-tjuw-ijo">https://meet.google.com/ioe-tjuw-ijo</a> <br> Please note that the meeting will be recorded as per Institute requirements. By joining the link you are giving your consent to the recording. DTSTART:20200828T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200827T120000Z UID:25dfbe0e0f4bfe368b6464ea97899481-92 DTSTAMP:19700101T120010Z DESCRIPTION:On building interactive and secure smart spaces URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/92/on-building-interactive-and-secure-smart-spaces/ SUMMARY:Please click on the following URL to join the talk. <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmU3ZGQzMzMtY2Q1MC00NzY0LWE2MzAtZGQ3NTIxNjgwZjFm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2282f39501-c5b2-4bfb-87c3-f17ca74c00b6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmU3ZGQzMzMtY2Q1MC00NzY0LWE2MzAtZGQ3NTIxNjgwZjFm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2282f39501-c5b2-4bfb-87c3-f17ca74c00b6%22%7d</a> <br> Absract: <br> The future of Internet-of-Things (IoT) demands seamless interaction between users and devices, where sensing interfaces blend into everyday objects. Toward realizing the vision of making sensing interfaces truly ubiquitous, we also need to make the future smart spaces secure. In this talk, I will present two of my works, which concern with these critical issues. Firstly, I will present RIO, a novel battery-free touch-sensing user interface (UI) primitive. With RIO, any surface can be turned into a touch-aware interactive surface by directly attaching RFID tags. RIO is built using impedance tracking: when a human finger touches the surface of an RFID tag, the impedance of the antenna changes. This change manifests as a variation in the phase of the RFID backscattered signal and is used by RIO to track fine-grained touch movement, over both off-the-shelf and custom-built tags. Secondly, I will present REVOLT, an end-to-end system to detect replay attacks on voice-first devices (e.g., Amazon Echo, Google Home, etc.) without requiring a user to wear any wearable device. This system has several distinct features: (i) it intelligently exploits the inherent differences between the spectral characteristics of the original and replayed voice signals, (ii) it exploits both acoustic and WiFi channels in tandem, (iii) it utilizes unique breathing rate extracted from WiFi signal while speaking to test the liveness of human voice. This novel technique of combining WiFi and voice modality yields low false positive and false negative when evaluated against a range of voice replay attacks. DTSTART:20200827T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200828T120000Z UID:ce99109e3c5db396ec3ba1299f65f70f-93 DTSTAMP:19700101T120016Z DESCRIPTION:Towards Secure and Efficient Realization of Pairing-Based Signatures from Static Assumptions URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/93/towards-secure-and-efficient-realization-of-pairing-based-signatures-from-static-assumptions/ SUMMARY:Bilinear pairing defined over elliptic curve group was first used to design novel cryptosystem in 2000. Since then a large number of cryptosystems has been proposed in pairing-based cryptography (PBC). The main tool for all such cryptosystems is the pairing map which can be defined on either composite or prime-order groups. Security of a public key cryptosystem is typically proved under some computational hardness assumption. PBC has witnessed a plenty of parameterized/interactive assumptions. However, it is well-known that such assumptions have several limitations. In this thesis we explore the question of security and efficiency of pairing-based signature schemes based on static assumptions. We have investigated the efficacy of the following approaches towards this goal: (i). frameworks for converting cryptosystems from composite to prime-order bilinear pairing setting, (ii). DejaQ framework, for removing dependency on parameterized assumption and (iii). dual-form signature (DFS) technique, for removing dependency on parameterized/interactive assumption. <br> First, we focus on the conversion framework. In 2005, Boneh et al. introduced a novel homomorphic encryption scheme using composite-order pairing setting. From then there are plenty of cryptosystems constructed in the composite-order pairing setting. However, it is well known that a composite-order pairing is significantly slower than its prime-order counterpart. This motivated Freeman to propose his conversion framework that converts some cryptographic protocols to the prime-order pairing setting. He formally defined certain properties called projecting and canceling, which are used in the protocol construction and/or in the security argument. Since then several frameworks have been proposed for conversion purpose. We revisit all the projecting frameworks and establish that Freemans framework is still optimal in the asymmetric pairing setting. We also present an alternative security proof for Seo-Cheons projecting and canceling framework under the static symmetric external Diffie-Hellman (SXDH) assumption, instead of the original tailor-made assumption. Next, we formalize the full-decomposition notion in the existing projecting frameworks and show that this notion is sufficient instead of the so-called translating property. Then, we abstract an unbalanced projecting framework in the asymmetric pairing setting that allows the pairing source groups to have different orders. As application, we convert the following schemes to the prime-order asymmetric pairing setting: Shacham-Waters ring signature, Boyen-Waters group signature and Meiklejohn et als round optimal blind signature. In their original construction, security of the above schemes requires both projecting and canceling properties in the composite-order symmetric pairing setting. We show that the framework for projecting and canceling is not necessary to instantiate these schemes. <br> Next, we focus on a set of parameterized assumptions called the BTA family. Such parameterized assumptions play a crucial role in the security of many novel pairing-based cryptosystems. However, they have some negative impact on efficiency at a concrete security level. A prominent approach to remove the dependency on parameterized assumption is the DejaQ framework of Chase et al. The DejaQ framework aims to establish that certain parameterized assumptions are implied by the subgroup hiding assumption in the composite-order pairing setting. Applying DejaQ framework to a family of assumptions is an important question, as it covers several parameterized assumptions which in turn cover more cryptographic protocols. However, the existing DejaQ framework could cover only Boyens Uber assumption family. Recently Ghadafi and Groth introduced the bilinear target assumption (BTA) family, that covers more parameterized assumptions including the Uber assumption family. We show that the parameterized assumptions that belong to the BTA family are reducible from the subgroup hiding assumption. In the process, we first suitably extend a property called parameter-hiding and then adapt the DejaQ proof technique on the parameterized assumptions that belong to the BTA family. <br> Finally, we focus on the applicability of the dual-form signature (DFS) technique on some pairing-based signatures. The DejaQ framework does not address the question of how to remove the dependency on interactive assumption. The DFS technique can be used for this purpose and it is applied directly in the security argument of the protocol. We use the DFS technique to prove the security of Abe et als structure-preserving signature, Boyen-Waters group signature and Pointcheval-Sanders rerandomizable signature (RRS) under some static assumptions. We also present an efficient construction of RRS scheme in the prime-order setting. Then, we use the proposed RRS scheme as a building block to construct a sequential aggregate signature (SeqAS) scheme with constant-size public key under the SXDH assumption. The performance of the proposed schemes is comparable to that of previous proposals based on some non-standard interactive assumptions. <br> <br> Link to join Microsoft Teams: <br> <a href="https://teams.microsoft.com/l/channel/19%3aa24e8f8dbe724c0486b5239881d7674c%40thread.tacv2/General?groupId=57f6f230-1dd8-4280-85ed-de7a55a7d936&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/channel/19%3aa24e8f8dbe724c0486b5239881d7674c%40thread.tacv2/General?groupId=57f6f230-1dd8-4280-85ed-de7a55a7d936&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20200828T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200901T120000Z UID:689f7824219fea4f69b015b159cc39fd-94 DTSTAMP:19700101T120014Z DESCRIPTION:Algorithms for Social Good in Online Platforms with Guarantees on Honest Participation and Fairness URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/94/algorithms-for-social-good-in-online-platforms-with-guarantees-on-honest-participation-and-fairness/ SUMMARY:Recent decades have seen a revolution in the way people communicate, buy products, learn new things, and share life experiences. This has spurred the growth of online platforms that enable users from all over the globe to buy/review/recommend products and services, ask questions and provide responses, participate in online learning, etc. <br> There are certain crucial requirements that are required to be satisfied by the online forums for ensuring their trustworthiness and sustainability. In this thesis, we are concerned with three of these requirements: social welfare maximization, honest participation by the users, and fairness in decision making. In particular, we address three contemporary problems in online platforms and obtain principled solutions that achieve social welfare maximization while satisfying honest participation and fairness of allocation. The three problems considered are set in the context of three different platforms: online review or Q&A forums, online discussion forums, and online search platforms. In each case, we develop an abstraction of the problem and solve it in its generality. <br> <br> Ballooning Multi-armed Bandits <br> In our first problem, we consider online platforms where the users are shown user generated content such as reviews on an e-commerce platform or answers on a Q&A platform. The number of reviews/answers increases over time. We seek to design an algorithm that quickly learns the best review/best answer and displays it prominently. We model this problem as a novel multi-armed bandit formulation (which we call ballooning bandits) in which the set of arms expands over time. We first show that when the number of arms grows linearly with time, one cannot achieve sub-linear regret. In a realistic special case, where the best answer is likely to arrive early enough, we prove that we can achieve optimal sublinear regret guarantee. We prove our results for best answer arrival time distributions that have sub-exponetal or sub-Pareto tails. <br> <br> Strategy-proof Allocation of Indivisible Goods with Fairness Guarantees <br> Second, we consider the problem of fairness in online search platforms. We view the sponsored ad-slots on these platforms as indivisible goods to be allocated in a fair manner among competing advertisers. We use envy-freeness up to one good (EF1) and maximin fair share (MMS) allocation as the fairness notions. The problem is to maximize the overall social welfare subject to these fairness constraints. We first prove under a single parameter setting that the problem of social welfare maximization under EF1 is NP-hard. We complement this result by showing that any EF1 allocation satisfies an 1/2-approximation guarantee and present an algorithm with a (1, 1/2) bi-criteria approximation guarantee. We finally show in a strategic setting that one can design a truthful mechanism with the proposed fair allocation. <br> <br> Coalition Resistant Credit Score Functions <br> In the third problem, we study manipulation in online discussion forums. We consider a specific but a common form of manipulation namely manipulation by coalition formation. We design a manipulation resistant credit scoring rule that assigns to each user a score such that forming a coalition is discouraged. In particular, we study the graph generated by the interactions on the platform and use community detection algorithms. We show that the community scores given by community detection algorithms that maximize modularity lead to a coalition resistant credit scoring rule. This in turn leads to sustainable discussion forums with honest participation from users, devoid of any coalitional manipulation. <br> <br> Microsoft Teams Meeting: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OGM2MGIyZGQtZjZkMi00YTAyLTg5MGItYTRiMGFkZjZmNmRm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2269dfbf19-0e41-447f-b7d4-700f9629e40e%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OGM2MGIyZGQtZjZkMi00YTAyLTg5MGItYTRiMGFkZjZmNmRm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2269dfbf19-0e41-447f-b7d4-700f9629e40e%22%7d</a> DTSTART:20200901T120000Z END:VEVENT BEGIN:VEVENT DTEND:20200915T120000Z UID:b1624732814df05be7f77688f1baed02-95 DTSTAMP:19700101T120015Z DESCRIPTION:Ludic design for Accessibility URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/95/ludic-design-for-accessibility/ SUMMARY:Technology solutions for accessibility have long been created using a narrow utilitarian lens, especially in the global south due to the multi-dimensional challenges and resource constraints. We propose an alternate design methodology called the Ludic Design for Accessibility (LDA) that puts play and playfulness at the center of all assistive technology design and use. We have been exploring the application of this methodology in developing solutions for diverse disabilities using a range of technologies. In this talk I will focus on the application of this methodology to the following challenge and the lessons being learnt: introducing digital skills and computational thinking to children in schools for the blind in India starting at grade 2. DTSTART:20200915T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201005T120000Z UID:7e13f8ca759979ae33409a373c26ca3f-96 DTSTAMP:19700101T120015Z DESCRIPTION:New Algorithmic and Hardness results in Learning, Error Correcting Codes, and Constraint Satisfaction Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/96/new-algorithmic-and-hardness-results-in-learning-error-correcting-codes-and-constraint-satisfaction-problems/ SUMMARY:Approximation is a natural way to deal with the intractability barrier that is inherent in many natural computational problems. However, it is often the case that the task of solving the approximation version of a problem is as hard as the exact version of the problem itself, which is the driving philosophy underlying the celebrated Probabilistically Checkable Proofs (PCP) theorem. In this thesis, we investigate the hardness of approximating several fundamental problems in Learning Theory, Coding Theory and Constraint Satisfaction Problems. <br> <br> -- Hardness of Learning of Halfspaces Using Polynomial Thresholds. <br> In the first part, we study the complexity of improper learning of halfspaces. Here we study the following question: given a distribution over point label pairs with the guarantee that there exists a halfspace that classifies almost all the point-label pairs (say, 99%) correctly, can we efficiently find a hypothesis from a (possibly different) hypothesis class, that classifies a non-trivial (say, 51%)-fraction of point-label pairs correctly. When the hypothesis class is the set of halfspaces itself (i.e, proper learning), it is well known that it is NP-Hard to weakly learn halfspaces. A natural next step is to consider hypothesis classes which generalize halfspaces (i.e, improper learning), with the hope that this relaxation would allow for efficient approximation algorithms. A candidate hypothesis class which is popular in this setting is the set of Polynomial Threshold Functions (PTFs). By inclusion, learning halfspaces using halfspaces using PTFs is at least as easy as learning halfspaces, which as is does not rule out the efficient weakly learning halfspaces using constant degree PTFs. However, the added twist of dealing with two distinct hypothesis classes makes the task of ruling out weak learning in this setting much more challenging. Consequently, previous results in this direction were only able to rule out weak learning using degree-2 PTFs. In this work, we show that it is NP-Hard to weakly learn halfspaces using any constant degree PTF. In fact, we rule out efficient weak learning of halfspaces using any function of constant number of constant degree PTFs. <br><br> -- Hardness of Learning DNFs using Halfspaces. <br> In the second part, we investigate the analogous question of weakly learning Disjunctive Normal Forms (DNFs). DNFs are a fundamental concept class, simply due to the fact that any boolean function admits a DNF representation, and the size of its DNF representation is often a useful measure of its complexity. This naturally motivates the question of agnostic learning of DNFs, where we are given a distribution over point-label pairs that is perfectly (or almost perfectly) classifiable by a small size DNF, one wishes to efficiently find a hypothesis that classifies a large fraction of these point-label pairs correctly. While analogous to the setting of halfspaces, it is NP-Hard to weakly learn constant terms DNFs with constant term DNFs, the similarities do not carry over to the setting of improper learning. In particular, it is folkore that constant term DNFs can be efficiently learnt using constant degree PTFs, thus motivating the question of whether DNFs can be weakly learnt using halfspaces? We show that it is NP-Hard to weakly learn constant term DNFs (or a noisy 1-term DNF) using any function of constant number of halfspaces. This simultaneously generalizes and strengthens previous known results on learning 2-term DNFs using t-term DNFs, learning noisy monomials using halfspaces, and learning intersection of two halfspaces with intersections of constant number of halfspaces. <br><br> -- Hardness of Solving Sparse Equations over Finite Fields. <br><br> Given a system of linear equations over a finite field that admits a k-sparse solution, can we find a Ck-sparse solution that satisfies all the equations? Can we efficiently find a equally sparse solution that satisfies a large fraction fraction of equations? These are natural questions that arise in the context of Learning Sparse Parities over finite fields and error correcting codes. We study these problems in the context of parameterized complexity, where the parameterization is in terms of the promised sparsity of the solution. Of particular importance here is the homogeneous variant of the former question, popularly termed as the EVEN-SET problem. Here one asks if there exists an Fixed Parameter Tractable (FPT) algorithm that can recover a approximately sparse non-trivial solution to a system of homogeneous linear equations. The fixed parameter tractability of even the exact version was a long standing open problem in parameterized complexity. We show that assuming Randomized Gap-ETH (Exponential Time Hypothesis), there are no FPT algorithms for the EVEN-SET problem, which give a constant factor approximation, for any constant. Our techniques also extend to rule out constant factor FPT approximation algorithms for the parameterized variant of the Shortest Vector Problem, which is a fundamental computational problems on lattices with widespread applications in cryptography. <br> <br> -- Vertex Deletion into Easy CSPs. <br> We consider the following question. Given a Constraint Satisfaction Problem (CSP) with the promise that deleting a few vertices makes it easy , are their efficient algorithms that can delete a relatively small number of vertices to find a large vertex induced easy sub-instance? This fairly general question models a wide range of combinatorial tasks such as deleting the smallest number of vertices to make a graph bipartite, or deleting the smallest number of variables to make a system of 2-variable equations satisfiable, to list a few. In particular, we study the UNIQUE-GAMES variant of the above problem -- known as STRONG-UNIQUE-GAMES -- where given a UNIQUE-GAMES instance with the promise that deleting a few vertices makes it fully satisfiable, the objective is to efficiently delete a relatively small subset of vertices to make the remaining instance fully satisfiable. In addition to being an interesting problem on its own, STRONG-UNIQUE-GAMES is a useful variant which is often easier to reduce from, and has been used to show tight UG inapproximability for several problems such as VERTEX-COVER, MAX-BI-CLIQUE, SCHEDULING-WITH-PRECEDENCE-CONSTRAINTS, etc. However the vertex deletion nature of this problem pushes its complexity into the realm of CSPs with global constraints, which are in general far less well understood in comparison to general Max CSPs. By exploiting a novel connection between STRONG-UNIQUE-GAMES and SMALL-SET-VERTEX-EXPANSION in graphs, we give new algorithmic and hardness results for STRONG UNIQUE GAMES, that are matching for constant alphabet sizes. In particular, this leads to tight bounds for the well studied ODD-CYCLE-TRANSVERSAL problem, which is the vertex deletion version of MAX-CUT. We also extend the algorithmic techniques used here to analogous variants of settings such as 3-COLORING and CSPs with Low Threshold Rank. <br><br> -- Testing Sparsity over Known and Unknown Bases. <br><br> In the final part, we take a detour from the framework of approximation algorithms and study the following question under the lens of property testing. Given a random sketch of a matrix, can we efficiently test whether it admits a sparse representation in terms of an unknown/known bases ? This can be thought of as the property testing version of the well studied Dictionary Learning problem, albeit, in a non-distributional setting. As expected, dictionary learning is a NP -Hard problem, and efficient algorithms with provable guarantees are known to exist only under distributional settings. Hence, apriori it is not self-evident whether an efficient algorithm exists for the testing version of the problem in the non-distributional setting, let alone a query efficient one. In this work, we give a simple testing algorithm for the above problem with some standard assumptions on A . Our tester is noise tolerant, and surprisingly, the number of sketches required by the tester is independent of the ambient dimension. The tester is based on gaussian width of sets, which is an intrinsic measure of complexity of geometric objects. The analysis of the tester combines tools from gaussian processes and high dimensional geometry. <br> <br> Microsoft Teams Meeting Link: <br> <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDdlY2UwZTMtZmM3OC00ODdjLWE2MzQtYmY0ZTA4MzJhMjg1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22dff3e365-caca-4573-a6d4-f5c40f3fd57b%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDdlY2UwZTMtZmM3OC00ODdjLWE2MzQtYmY0ZTA4MzJhMjg1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22dff3e365-caca-4573-a6d4-f5c40f3fd57b%22%7d</a> DTSTART:20201005T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201015T120000Z UID:fc73000c6170c4ca036d86d0297b5742-97 DTSTAMP:19700101T120018Z DESCRIPTION:Accelerator-level Parallelism URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/97/accelerator-level-parallelism/ SUMMARY:Computer system performance has improved due to creatively using more transistors (Moore's Law) in parallel via bit-, instruction-, thread-, and data-level parallelism. With the slowing of technology scaling, a way to further improve computer system performance under energy constraints is to employ hardware accelerators. Each accelerator is a hardware component that executes a targeted computation class faster and usually with (much) less energy. Already today, many chips in mobile, edge and cloud computing concurrently employ multiple accelerators in what we call accelerator-level parallelism (ALP). <br> This talk develops our hypothesis that ALP will spread to computer systems more broadly. ALP is a promising way to dramatically improve power-performance to enable broad, future use of deep Al, virtual reality, self-driving cars, etc. To this end, we review past parallelism levels and the ALP already present in mobile systems on a chip (SoCs). We then aid understanding of ALP with the Gables model and charge computer science researchers to develop better ALP "best practices" for: targeting accelerators, managing accelerator concurrency, choreographing inter-accelerator communication, and productively programming accelerators. This joint work with Vijay Janapa Reddi of Harvard. See also: https://www.sigarch.org/accelerator-level-parallelism/. DTSTART:20201015T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201009T120000Z UID:5a42038243b5eccac06abfbb93ecef8b-98 DTSTAMP:19700101T120011Z DESCRIPTION:Making Microarchitecture-impossible Possible for Performance and Security URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/98/making-microarchitecture-impossible-possible-for-performance-and-security/ SUMMARY:Single-thread performance is the key to the performance of desktops, laptops, servers, and handheld devices. However, in recent years, the industry trend shows that improvements in single-thread performance are starting to saturate as Moore's law comes to an end, making it impossible to improve performance further. The talk will focus on making the impossible, possible through a bouquet of hardware prefetchers. General purpose CPUs usually suffer from a backend bottleneck, where data prefetchers come into the picture to hide the memory latency. Recent data prefetchers use monolithic and gigantic structures demanding 100s of KBs to improve application performance, making it impossible for the industry to adopt these ideas. Even after 30 years of prefetching research, commercial processors still use 30 years-old prefetchers. The talk will highlight a bouquet-based tiny prefetcher that outperforms all the data prefetchers proposed in the last 30 years. <br> Finally, the talk will provide a top-level view on the security side of microarchitecture, highlighting some of the offensive and defensive sides that seem impossible. On the offensive side, the talk will highlight the DABANGG cache attack that works on noisy systems, and on the defensive side, the talk will touch upon a lightweight and secure memory hierarchy for Trusted Execution Environments (TEEs). DTSTART:20201009T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201012T120000Z UID:ae07170fdf9c32907741a276b2fb5122-99 DTSTAMP:19700101T120015Z DESCRIPTION:Algorithms for Fair Decision Making: Provable Guarantees and Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/99/algorithms-for-fair-decision-making-provable-guarantees-and-applications/ SUMMARY:The topic of fair allocation of indivisible items has received significant attention because of its applicability in several real-world settings. This has led to a vast body of work focusing on defining appropriate fairness notions, providing existence guarantees, and handling computational issues in obtaining such fair allocations. This thesis addresses open questions in the space of fair allocation of indivisible items. We study constrained versions of the fair allocation problem, namely cardinality constraints and matroid constraints. These constrained settings generalize the unconstrained formulation and we are the first to study these constructs. We establish the existence of well-studied fairness notions (such as EF1 and MMS) under the constrained settings, and we design methods that provide an algorithmic anchor to these existence results. Moreover, we define strictly stronger notions of fairness and provide algorithms for obtaining these stronger fairness guarantees. Finally, we investigate fairness in diverse application scenarios, such as recommendation systems and classification problems. The key novelty involves providing solutions to such applications through the lens of fair allocation. <br> Fair Allocation under Cardinality Constraints <br> We investigate the problem of fairly allocating goods under cardinality constraints and additive valuations. In this setting, the set of goods are categorized, and an upper limit is imposed on the number of goods allocated to any agent from a particular category. The objective is to find an allocation that satisfies the given cardinality constraints as well as a fairness constraint. We design an efficient algorithm that computes an envy-free up to one good (EF1) allocation. Additionally, this algorithm outputs an exact maximin share (MMS) allocation when the valuations are binary. We also show that the constrained fair allocation problem with additive valuations reduces to an unconstrained fair allocation problem with submodular valuations. This allows us to guarantee 1/3-approximate maximin share (1/3-MMS) allocations under cardinality constraints. <br> Fair Allocation under Matroid Constraints <br> We consider the fair allocation problem under more general constraints. Here, each allocated bundle, in addition to the fairness criterion, needs to satisfy the independence criterion specified by a matroid. We establish that MMS exists in this setting when the valuations are identical. We provide an algorithm that efficiently computes an EF1 allocation under identical laminar matroid. The algorithm initializes an allocation by computing a matroid feasible partition of goods (using a method proposed by Gabow and Westermann) and then iteratively reallocates goods between the bundles till an EF1 allocation is obtained. Our reallocation strategy maintains matroid feasibility at each iteration (using an extension of the strong basis exchange lemma) and also ensures polynomial time convergence. <br> Stronger Notions of Fairness <br> We define two novel fairness notions, namely envy-free up to one less preferred good (EFL) and groupwise maximin share (GMMS). We show that these fairness notions are better, in terms of social welfare, compared to EF1 and MMS, respectively. We provide a scale of fairness to establish how these new fairness notions fit in the hierarchy of existing notions. We provide an algorithm that outputs an EFL and ½-GMMS allocations under the unconstrained setting. We also show that exact GMMS allocations are guaranteed to exist when the valuations of the agents are either binary or identical. We empirically show that GMMS allocations exist when the valuations are drawn from Gaussian and Uniform distributions. These results highlight that, for unconstrained settings, we do not fall short on generic existence results by strengthening the existing fairness notions. <br> Application to Two-Sided Fair Recommendation Systems <br> We investigate the problem of fair recommendation in two-sided online platforms, such as Amazon, Netflix, and Spotify, consisting of customers on one side and producers on the other. These services have typically focused only on maximizing customer satisfaction by tailoring the recommendations according to the preferences of individual customers, which may be detrimental for the producers. We consider fairness issues that span both customers and producers. Our approach involves a mapping of the fair recommendation problem to a constrained version of the fair allocation problem. Our proposed algorithm guarantees at least MMS exposure for most of the producers and EF1 fairness for every customer. We establish theoretical guarantees and provide empirical evidence through extensive evaluations on real-world datasets. <br> Application to Classification Problems under Prior Probability Shifts <br> We consider the problem of fair classification under prior probability shifts, which is a kind of distributional change occurring between the training and test datasets. Such shifts can be observed in the yearly records of several real-world datasets, such as COMPAS. If unaccounted for, such shifts can cause the predictions to become unfair towards specific population sub-groups. We define a fairness notion, called proportional equality (PE) which is motivated by solution concepts from the fair allocation literature, and accounts for prior probability shifts. We develop an algorithm CAPE that uses prevalence estimation techniques, sampling and an ensemble of classifiers to ensure fair predictions. We evaluate the performance of CAPE on real-world datasets and compare its performance with state-of-the-art fair algorithms. Our findings indicate that CAPE ensures PE-fair predictions, with low compromise on other performance metrics. <br> <br> Microsoft Teams link : <br> <a href="https://teams.microsoft.com/l/team/19%3a55c4dc3c4baa4347bff91658e32050ff%40thread.tacv2/conversations?groupId=9c444335-6a84-4c98-8a99-b93aa0744052&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/team/19%3a55c4dc3c4baa4347bff91658e32050ff%40thread.tacv2/conversations?groupId=9c444335-6a84-4c98-8a99-b93aa0744052&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20201012T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201014T120000Z UID:7cbac09cbf7c673451183d46a0e481b1-100 DTSTAMP:19700101T120011Z DESCRIPTION:Decision Making under Uncertainty : Reinforcement Learning Algorithms and Applications in Cloud Computing, Crowdsourcing and Predictive Analytics URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/100/decision-making-under-uncertainty-reinforcement-learning-algorithms-and-applications-in-cloud-computing-crowdsourcing-and-predictive-analytics/ SUMMARY:In this thesis, we study both theoretical and practical aspects of decision making, with a focus on reinforcement learning based methods. Reinforcement learning (RL) is a form of semi-supervised learning in which the agent learns the decision making strategy by interacting with its environment. We develop novel reinforcement learning algorithms and study decision problems in the domains of cloud computing, crowdsourcing and predictive analytics. <br> In the first part of the thesis, we develop a model free reinforcement learning algorithm with faster convergence named Generalized Speedy Q-learning and analyze its finite time performance. This algorithm integrates ideas from the well-known Speedy Q-learning algorithm and the generalized Bellman equation to derive a simple and efficient update rule such that its finite time bound is better than that of Speedy Q-learning for MDPs with a special structure. Further, we extend our algorithm to deal with large state and action spaces by using function approximation. <br> Extending the idea in the above algorithm, we develop a novel Deep Reinforcement Learning algorithm by combining the technique of successive over-relaxation with Deep Q-networks. The new algorithm, named SOR-DQN, uses modified targets in the DQN framework with the aim of accelerating training. We study the application of SOR-DQN in the problem of auto-scaling resources for cloud applications, for which existing algorithms suffer from issues such as slow convergence, poor performance during the training phase and non-scalability. <br> Next, we consider an interesting research problem in the domain of crowdsourcing - that of efficiently allocating a fixed budget among a set of tasks with varying difficulty levels. Further, the assignment of tasks to workers with different skill levels is tackled. This problem is modeled in the RL framework and an approximate solution is proposed to deal with the exploding state space. <br> We also study the following problem in predictive analytics : predicting the future values of system parameters well in advance for a large-scale software or industrial system, which is important for avoiding disruptions. An equally challenging and useful exercise is to identify the 'important' parameters and optimize them in order to attain good system performance. In addition to devising an end-to-end solution for the problem, we present a case study on a large-scale enterprise system to validate the effectiveness of the proposed approach. <br> Link to the Online Thesis Defense: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3a6e490d72f0b34eac8e690f8daac4a00a%40thread.tacv2/1602060940579?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2286976192-e64f-4e50-ae9f-8a79b451c8f8%22%7d">https://teams.microsoft.com/l/meetup-join/19%3a6e490d72f0b34eac8e690f8daac4a00a%40thread.tacv2/1602060940579?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2286976192-e64f-4e50-ae9f-8a79b451c8f8%22%7d</a> DTSTART:20201014T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201012T120000Z UID:0269111a8e94c9d3a357b1c4e50b83e0-101 DTSTAMP:19700101T120018Z DESCRIPTION:Statistics, computation and adaptation in offline and online learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/101/statistics-computation-and-adaptation-in-offline-and-online-learning/ SUMMARY:In this talk, I will address the statistical, computational and adaptive aspects of learning theory; both in offline (batch) as well as in online settings. The first one-third of the talk deals with a computationally efficient and statistically sound Alternating Minimization (AM) algorithm (often called hard EM), typically used to solve non-convex problems. In particular, we apply AM to a classical non-convex problem, namely max-affine regression. Max-affine regression can be thought of as a generalization of the (real) Phase Retrieval problem, and closely resembles the canonical problem of convex regression in non-parametric statistics. In the next segment of the talk, I characterize the (exact) convergence speed of the AM algorithm. In particular, a super-linear convergence of AM is (theoretically) proved, resolving a long-standing (1995) conjecture of Lei Xu and Micheal I. Jordan. The final part of the talk deals with adaptation, in a non-trivial online (bandit) setting. I will talk about my recent works on model selection in contextual bandits, which partially solves an open problem of COLT 2020. DTSTART:20201012T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201020T120000Z UID:e2f4ccac0fce8c99e54fa22edd47912b-102 DTSTAMP:19700101T120011Z DESCRIPTION:Algorithms for Stochastic Optimization, Statistical Estimation and Markov Decision Processes URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/102/algorithms-for-stochastic-optimization-statistical-estimation-and-markov-decision-processes/ SUMMARY:Stochastic approximation deals with the problem of finding zeros of a function expressed as an expectation of a random variable. In this thesis we study and analyze convergence of stochastic approximation algorithms in the context of optimization under uncertainty, statistical estimation and Reinforcement Learning. Moreover, we also explore second order methods in the context of Markov Decision Processes. <br> First, we consider Stochastic Optimization (SO) problem where we optimize an objective function in the presence of noise. A prominent algorithm in SO namely Random Direction Kiefer-Wolfowitz (RDKW) solves the problem by obtaining noisy gradient estimate by randomly perturbing all the parameters simultaneously. In this thesis, we characterize the class of deterministic perturbation sequences that can be utilized in the RDKW algorithm. Using our characterization, we propose a construction of a deterministic perturbation sequence that has the least possible cycle length among all deterministic perturbations. We establish the convergence of the RDKW algorithm for the generalized class of deterministic perturbations. <br> In statistical estimation, one of the popular measures of central tendency that provides better representation and interesting insights of the data compared to the other measures like mean and median is the metric mode. In the second part of our thesis, we provide a computationally effective, on-line iterative algorithm that estimates the mode of a unimodal smooth density given only the samples generated from the density. Asymptotic convergence of the proposed algorithm using stochastic approximation techniques is provided. We also prove the stability of the mode estimates by utilizing the concept of regularization. <br> In the third part of our thesis, we propose Successive Over-Relaxation (SOR) Q-learning. In a discounted reward Markov Decision Process (MDP), the objective is to find the optimal value function. We first derive a modified fixed point iteration for SOR Q-values and utilize stochastic approximation to derive a learning algorithm to compute the optimal value function and an optimal policy. We then prove the almost sure convergence of the SOR Q-learning to SOR Q-values. Finally, through numerical experiments, we demonstrate that SOR Q-learning is faster compared to the standard Q-learning algorithm in the literature. <br> In the fourth part of our thesis, we explore second-order methods in MDPs. We propose a second order value iteration procedure that is obtained by applying the Newton-Raphson method to the successive relaxation value iteration scheme. We prove the global convergence of our algorithm to the optimal solution asymptotically and show the second order convergence. Through experiments, we show the effectiveness of our proposed approach. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWVlZTEwZTQtYjQ5OC00NGZlLWE2NzAtYTkxOGJkZjBjZjJi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22733c69c9-0556-4626-89a4-cd9760e97080%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWVlZTEwZTQtYjQ5OC00NGZlLWE2NzAtYTkxOGJkZjBjZjJi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22733c69c9-0556-4626-89a4-cd9760e97080%22%7d</a> DTSTART:20201020T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201102T120000Z UID:4566bb54f51972ed044b0b617110d1e8-104 DTSTAMP:19700101T120016Z DESCRIPTION:Robust Algorithms for recovering planted structures in Semi-random instances URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/104/robust-algorithms-for-recovering-planted-structures-in-semi-random-instances/ SUMMARY:In this work, we design algorithms for two fundamentally important and classical graph problems in the planted setting. Both of these problems are NP-hard and the bounds known from the algorithmic front are either not fully understood, or not much progress can be made because of tight lower bounds. Thus it is natural to consider semi-random models for these problems. These models are inspired from the seminal paper of Feige and Killian [FK01] and have been studied in numerous follow-up works with the latest ones by Steinhardt, and McKenzie et al. [Ste17, MMT20]. The construction of our instance starts with an empty graph, then an arbitrary set of vertices (S) is chosen and either a dense graph or a clique (or an independent set) is planted on it, the subgraph on S x V-S is a random graph, while the subgraph on V-S might be a random, arbitrary, or some special graph (depending on the model). Our algorithms are based on rounding semi-definite programs and our primary focus is on recovering (completely or partially) the planted structure (S) with high probability (over the randomness of the input). We give algorithms that exploit the geometry of the corresponding vectors (from the SDP) and are easy to design/analyse. <br> The two problems which we study are: <br> -- Densest k-Subgraph/Clique<br> Given an undirected graph G, the Densest k-Subgraph problem (DkS) asks to compute a set S subseteq V of cardinality k such that the weight of edges inside S is maximized. This is a fundamental NP-hard problem whose approximability, inspite of many decades of research, is yet to be settled. There is a significant gap between the best known worst-case approximation factor of this problem [BCC+10] and the hardness of approximation for it (assuming the Exponential Time Hypothesis) [Man17]. We ask what are some easier instances of this problem? We propose some natural semi-random models of instances with a planted dense subgraph, and study approximation algorithms for computing the densest subgraph in them. There are many such random and semi-random models which have been studied in the literature [BCC+10, Ame15, HWX16, BA19 etc.]. <br> <br> -- Independent Set in Hypergraphs<br> The independent set problem in graphs poses the following question : given a graph, and a subset of vertices such that any two vertices of the set do not have an edge between them. The maximization version of this problem features as one of the Karp's original twenty-one NP-complete problems ([Kar72], the clique problem instead of its complement, the independent set problem). The independent set problem is relatively well understood and by the famous result of HÃ¥stad [HÃ¥s99], the lower and upper bounds of this problem are tight. Hypergraphs are a natural extension of graphs, where each edge is formed across a tuple of distinct vertices. For a graph, each tuple has a size, two. To the best of our knowledge, semi-random models on hypergraphs have not been studied so far. Studying classical problems like these on hypergraphs is naturally of theoretical as well as practical interest. We study the algorithmic problems studied in McKenzie et al. [MMT20] and develop algorithms for them in the case of hypergraphs. <br> Note : Both of these e-prints are available online on arXiv<br> DTSTART:20201102T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201028T120000Z UID:293ca5def88cc877de706e186ce3e6c6-105 DTSTAMP:19700101T120009Z DESCRIPTION:Representing Networks: Centrality, Node Embeddings, Community Outliers and Graph Representation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/105/representing-networks-centrality-node-embeddings-community-outliers-and-graph-representation/ SUMMARY:Networks are ubiquitous. We start our technical work in this thesis by exploring the classical concept of node centrality (also known as influence measure) in information networks. Like clustering, node centrality is also an ill-posed problem. There exist several heuristics and algorithms to compute the centrality of a node in a graph, but there is no formal definition of centrality available in the network science literature. Lately, researchers have proposed axiomatic frameworks for the centrality of a node in a network. However, these existing formal frameworks are not generic in nature in terms of characterizing the space of influence measures in complex networks. In this work, we propose a set of six axioms in order to capture most of the intrinsic properties that any influence measure ideally should satisfy. We also characterize existing measures of centrality with respect to this framework. Next, we focus more on the representation learning on networks. Network embedding is required as real life networks are large, extremely sparse and discrete in nature. We investigate the problem of unsupervised node representation in attributed networks through informative random walk. Edges are also useful for various downstream network mining tasks, but most of the existing homogeneous network representation learning approaches focus on embedding the nodes of a graph. So, we propose a novel unsupervised algorithm to embed the edges of a network, through the application of the classical concept of line graph of a network. The optimization framework of edge embedding connects to the concept of node centrality in the representation learning framework. Also, we conduct research on attributed hypergraphs. We propose a novel hypergraph neural network to represent and classify hypernodes. Outlier analysis is another important problem in network science. All the real-world networks contain outlier nodes to some extent. Empirically we have shown that outliers can affect the quality of network embedding if not handled properly. So, we integrate the process of network embedding and outlier detection into a single framework. In this research thread, we first propose a matrix factorization based approach which minimizes the effect of outlier nodes in the framework of attributed network embedding. Next, we propose two neural network architectures, based on L2 regularization and adversarial training respectively, to minimize the effect of outliers on node embedding of an attributed network. Further, extending the concept of support vector data description, we propose a novel algorithm which integrates node embedding, community detection and outlier detection into a single optimization framework by exploiting the link structure of a graph. In the last part of the thesis, we focus on graph level representation and tasks. First, we propose a supervised graph neural network based algorithm with hierarchical pooling strategy to classify a graph from a set of graphs. Next, we propose a novel GNN based algorithm for the unsupervised representation of a graph from a set of graphs, so that similar graphs are represented closely in the embedding space and dissimilar graphs are separated away. Link to Microsoft Teams: https://teams.microsoft.com/l/channel/19%3a8c5f66cd10ea447eaa90e6c358e7e4d8%40thread.tacv2/General?groupId=3e16bdae-e641-4e8e-83e3-3719a65cdff2&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476 DTSTART:20201028T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201028T120000Z UID:ec87ab7f7a72b5e82d4d248a34f3a753-106 DTSTAMP:19700101T120011Z DESCRIPTION:Embedding Networks: Node and Graph Level Representations URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/106/embedding-networks-node-and-graph-level-representations/ SUMMARY:Network representation learning is essential to carry out various network analysis tasks. Graphs are the most suitable structures to represent real-world relational data such as social networks and molecular structures. In this thesis work, we focus on learning representations of the nodes and the entire graphs. Graph neural networks gained significant attention for graph representation and classification. <br> For graph classification, existing pooling approaches do not consider both the region-based and the graph level importance of the nodes together. We address this issue in the first part of the thesis by proposing a novel graph pooling layer R2POOL, which retains the most informative nodes for the next coarser version of the graph. Further, we integrate R2POOL with our multi-level prediction and branch training strategies to learn graph representations and to further enhance the model's capability for graph classification. <br> Moreover, the attention mechanisms on graphs improve the performance of graph neural networks. Typically, it helps to identify a neighbor node which plays a more important role in determining the label of the node under consideration. But in the real-world situation, a particular subset of nodes together may be significant. In the second part of the thesis, we address this problem and introduce the concept of subgraph attention for graphs. To show the efficiency of this scheme, we use subgraph attention for node classification. <br> Additionally, the hierarchical graph pooling is promising in graph literature. But, not all the graphs at different levels play an equal role in graph classification. Towards this end, we propose a novel algorithm called SubGattPool, which jointly learns the subgraph attention and employs two different attention mechanisms to find the important nodes in a hierarchy and the individual hierarchies in a graph for embedding and classifying the graphs given a collection of graphs. Improved performance over the state-of-the-art on both the real world and synthetic graphs for node and graph classification shows the efficiency of the proposed algorithms. <br> Link to Microsoft Teams: <br> <a href="https://teams.microsoft.com/l/team/19%3a4c60cfad6fdf4fbab81dd8df3991949a%40thread.tacv2/conversations?groupId=1684647f-67d5-4b7b-9313-61e8a6900b61&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/team/19%3a4c60cfad6fdf4fbab81dd8df3991949a%40thread.tacv2/conversations?groupId=1684647f-67d5-4b7b-9313-61e8a6900b61&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20201028T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201113T120000Z UID:704971bb19247d199d0b9cbb374d4bd0-107 DTSTAMP:19700101T120011Z DESCRIPTION:An O(m^{4/3+o(1)}) algorithm for exact unit-capacitated max flow URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/107/an-om4-3o1-algorithm-for-exact-unit-capacitated-max-flow/ SUMMARY:In this talk, I will present an algorithm for computing s-t max flows in O(m^{4/3+o(1)}) time in unit capacity graphs. The algorithm is inspired by potential reduction Interior Point Methods for linear programming. Instead of using scaled gradient/Newton steps of a potential function, we consider minimizing the potential function exactly and show how this allows us to directly find a centered point efficiently. Then using the weighted central path framework of Madry and Liu-Sidford, we show that our steps also benefit from maximizing weights carefully, which can be efficiently solved using work of Kyng et al., which allows us to get the desired runtime. <br> <br> Link to Microsoft Teams meeting: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MjY0NmJhNWMtMDdkNC00ZWIxLTgzMWMtMjExNjZiYTA3YmE2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22febce5be-d427-430f-b9d5-0bab6673a9ed%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MjY0NmJhNWMtMDdkNC00ZWIxLTgzMWMtMjExNjZiYTA3YmE2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22febce5be-d427-430f-b9d5-0bab6673a9ed%22%7d</a> DTSTART:20201113T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201201T120000Z UID:619ad65158c0f967e851e5cd0f01b757-108 DTSTAMP:19700101T120014Z DESCRIPTION:Algorithms for Challenges to Practical Reinforcement Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/108/algorithms-for-challenges-to-practical-reinforcement-learning/ SUMMARY:Reinforcement learning (RL) in real world applications faces major hurdles - the foremost being safety of the physical system controlled by the learning agent and the varying environment conditions in which the autonomous agent functions. A RL agent learns to control a system by exploring available actions. In some operating states, when the RL agent exercises an exploratory action, the system may enter unsafe operation, which can lead to safety hazards both for the system as well as for humans supervising the system. RL algorithms thus need to respect these safety constraints and must do so with limited available information. Additionally, RL autonomous agents learn optimal decisions in the presence of a stationary environment. However, the stationary assumption on the environment is very restrictive. In many real world problems like traffic signal control, robotic applications, etc., one often encounters situations with non-stationary environments, and in these scenarios, RL algorithms yield sub-optimal decisions. <br> This talk describes our algorithmic solutions to the challenges of safety and non-stationary environmental conditions in RL. In order to handle safety restrictions and facilitate safe exploration during learning, we propose a cross-entropy method based sample efficient learning algorithm. This algorithm is developed based on constrained optimization framework and utilizes limited information for the learning of feasible policies. Also, during the learning iterations, the exploration is guided in a manner that minimizes safety violations. <br> The goal of the second algorithm is to find a good policy for control when the latent model of the environment changes with time. To achieve this, the algorithm leverages a change point detection algorithm to monitor change in the statistics of the environment. The results from this statistical algorithm are used to reset learning of policies and efficiently control an underlying system. Both the proposed algorithms are tested numerically on benchmark problems in RL. <br> The second part of this talk will focus on application of RL to obstacle avoidance problem in UAV quadrotor. Obstacle avoidance in quadrotor aerial vehicle navigation brings in additional challenges in comparison to ground vehicles. This is because, an aerial vehicle needs to navigate across more types of obstacles - for e.g., objects like decorative items, furnishings, ceiling fans, sign-boards, tree branches, etc., are also potential obstacles for a quadrotor aerial vehicle. Thus, methods of obstacle avoidance developed for ground robots are clearly inadequate for UAV navigation. Our proposed method utilizes the relevant temporal information available from the ambient surroundings for this problem. This information is a sequence of monocular camera images collected by the quadrotor aerial vehicle. Our method adapts attention based deep Q networks combined with generative adversarial networks for this application. It improves efficiency of learning by inferring quadrotor maneuver decisions from temporal information of the ambient surroundings. <br> <br> Microsoft Teams Meeting Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_M2E3MmRlZjktY2Y0Mi00ZmZhLTljMDgtYTg4NThiYWJhZmU3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%221b0586c2-1488-4f8f-ab3c-d2e61940254c%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_M2E3MmRlZjktY2Y0Mi00ZmZhLTljMDgtYTg4NThiYWJhZmU3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%221b0586c2-1488-4f8f-ab3c-d2e61940254c%22%7d</a> DTSTART:20201201T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201203T120000Z UID:dd1048862a51bde50d7424a3ae71619a-109 DTSTAMP:19700101T120010Z DESCRIPTION:Towards Secure and Efficient Realization of Pairing-Based Signatures from Static Assumptions. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/109/towards-secure-and-efficient-realization-of-pairing-based-signatures-from-static-assumptions/ SUMMARY:Bilinear pairing defined over elliptic curve group was first used to design novel cryptosystem in 2000. Since then a large number of cryptosystems has been proposed in pairing-based cryptography (PBC). The main tool for all such cryptosystems is the pairing map which can be defined on either composite or prime-order groups. Security of a public key cryptosystem is typically proved under some computational hardness assumption. PBC has witnessed a plenty of parameterized/interactive assumptions. However, it is well-known that such assumptions have several limitations. In this thesis we explore the question of security and efficiency of pairing-based signature schemes based on static assumptions. We have investigated the efficacy of the following approaches towards this goal: (i). frameworks for converting cryptosystems from composite to prime-order bilinear pairing setting, (ii). DejaQ framework, for removing dependency on parameterized assumption and (iii). dual-form signature (DFS) technique, for removing dependency on parameterized/interactive assumption. <br> First, we focus on the conversion framework. In 2005, Boneh et al. introduced a novel homomorphic encryption scheme using composite-order pairing setting. From then there are plenty of cryptosystems constructed in the composite-order pairing setting. However, it is well known that a composite-order pairing is significantly slower than its prime-order counterpart. This motivated Freeman to propose his conversion framework that converts some cryptographic protocols to the prime-order pairing setting. He formally defined certain properties called projecting and canceling, which are used in the protocol construction and/or in the security argument. Since then several frameworks have been proposed for conversion purpose. We revisit all the projecting frameworks and establish that Freemans framework is still optimal in the asymmetric pairing setting. We also present an alternative security proof for Seo-Cheons projecting and canceling framework under the static symmetric external Diffie-Hellman (SXDH) assumption, instead of the original tailor-made assumption. Next, we formalize the full-decomposition notion in the existing projecting frameworks and show that this notion is sufficient instead of the so-called translating property. Then, we abstract an unbalanced projecting framework in the asymmetric pairing setting that allows the pairing source groups to have different orders. As application, we convert the following schemes to the prime-order asymmetric pairing setting: Shacham-Waters ring signature, Boyen-Waters group signature and Meiklejohn et al.s round optimal blind signature. In their original construction, security of the above schemes requires both projecting and canceling properties in the composite-order symmetric pairing setting. We show that the framework for projecting and canceling is not necessary to instantiate these schemes. <br> Next, we focus on a set of parameterized assumptions called the BTA family. Such parameterized assumptions play a crucial role in the security of many novel pairing-based cryptosystems. However, they have some negative impact on efficiency at a concrete security level. A prominent approach to remove the dependency on parameterized assumption is the DejaQ framework of Chase et al. The DejaQ framework aims to establish that certain parameterized assumptions are implied by the subgroup hiding assumption in the composite-order pairing setting. Applying DejaQ framework to a family of assumptions is an important question, as it covers several parameterized assumptions which in turn cover more cryptographic protocols. However, the existing DejaQ framework could cover only Boyens Uber assumption family. Recently Ghadafi and Groth introduced the bilinear target assumption (BTA) family, that covers more parameterized assumptions including the Uber assumption family. We show that the parameterized assumptions that belong to the BTA family are reducible from the subgroup hiding assumption. In the process, we first suitably extend a property called parameter-hiding and then adapt the DejaQ proof technique on the parameterized assumptions that belong to the BTA family. <br> Finally, we focus on the applicability of the dual-form signature (DFS) technique on some pairing-based signatures. The DejaQ framework does not address the question of how to remove the dependency on interactive assumption. The DFS technique can be used for this purpose and it is applied directly in the security argument of the protocol. We use the DFS technique to prove the security of Abe et al.s structure-preserving signature, Boyen-Waters group signature and Pointcheval-Sanders rerandomizable signature (RRS) under some static assumptions. We also present an efficient construction of RRS scheme in the prime-order setting. Then, we use the proposed RRS scheme as a building block to construct a sequential aggregate signature (SeqAS) scheme with constant-size public key under the SXDH assumption. The performance of the proposed schemes is comparable to that of previous proposals based on some non-standard interactive assumptions. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/team/19%3ab2e4009d9ba64dd1a4608e1e4d71634f%40thread.tacv2/conversations?groupId=01bc59f0-ab93-44a2-8bab-93177163507e&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/team/19%3ab2e4009d9ba64dd1a4608e1e4d71634f%40thread.tacv2/conversations?groupId=01bc59f0-ab93-44a2-8bab-93177163507e&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20201203T120000Z END:VEVENT BEGIN:VEVENT DTEND:20201208T120000Z UID:4d74fbc9b9777342ffcd010a55a42428-110 DTSTAMP:19700101T120010Z DESCRIPTION:Hypergraph Network Models: Learning, Prediction, and Representation in the Presence of Higher-Order Relations URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/110/hypergraph-network-models-learning-prediction-and-representation-in-the-presence-of-higher-order-relations/ SUMMARY:The very thought about relating objects makes us assume the relation would be pairwise, and not of a higher-order -- involving possibly more than two of them at a time. Yet in reality, higher-order relations do exist and are spread across multiple domains: medical science (e.g., co-existing diseases/symptoms), pharmacology (e.g., reacting chemicals), bibliometrics (e.g., collaborating researchers), the film industry (e.g., cast/crew), human resource (e.g., a team), social sciences (e.g., negotiating/conflicting nations), and so on. Since a collection of intersecting higher-order relations lose context when represented by a graph, "hypergraphs" -- graph-like structures that allow edges (called hyperedges/hyperlinks) spanning possibly more than two nodes -- capture them better. In a quest to better understand such relations, in this thesis we focus on solving a few network-science oriented problems involving hypergraphs. <br> "Hypergraphs and Pairwise Links": In the first of three broad parts, we study the behavior of usual graph-oriented networks that have an otherwise-ignored hypergraph underpinning. We particularly establish the skewness a hypergraph introduces into its induced graphs, and the effect of these biases on the structure and evaluation of the well-known problem of link prediction in networks. We find that an underlying hypergraph structure makes popular heuristics such as common-neighbors overestimate their ability to predict links. Gathering enough evidence -- both theoretical and empirical -- to support the need to reestablish the evaluations of link prediction algorithms on hypergraph-derived networks, we propose adjustments that essentially undo the undesired effects of hypergraphs in performance scores. Motivated by this observation, we extend graph-based structural node similarity measures to cater to hypergraphs (although still, for similarity between pairs of nodes). To be specific, we first establish mathematical transformations that could transfer any graph-structure-based notion of similarity between node pairs to a hypergraph-structure-based one. Using exhaustive combinations of the newly established scores with the existing ones, we could show improvements in the performance of both structural as well as temporal link prediction. <br> "Predicting Higher-order Relations": For the second part of our thesis, we turn our attention towards a more central problem in hypergraphs -- the hyperlink/hyperedge prediction problem. It simply refers to developing models to predict the occurrence of missing or future hyperedges. We first study the effect of negative sampling (sampling the negative class) -- an exercise performed due to the extreme intractability of the set of all non-hyperlinks, also known as the class imbalance problem -- on hyperlink prediction, which has never been studied in the past. Since we observe hyperlink prediction algorithms performing differently under different negative sampling techniques, our experiments help the seemingly unimportant procedure gain some significance. Moreover, we contribute towards two benchmark negative sampling algorithms that would help standardize the step. Moving on from the negative sampling problem to predicting hyperlinks themselves, we work on two different approaches: a clique-closure based approach, and a sub-higher-order oriented one. While in the former, we develop and successfully test the clique-closure hypothesis -- that hyperlinks mostly form from cliques or near-cliques -- and are able to utilize it for hyperlink prediction via matrix completion (C3MM), the latter approach works on a novel information flow model in hypergraphs. More precisely, we introduce the concept of sub-hyperedges to capture the sub-higher-order notion in relations, and utilize an attention-based neural network model called SHONeN focusing on sub-hyperedges of a hyperedge. Owing to SHONeNs computational complexity, we propose a sub-optimal heuristic that is able to perform better than its baselines on the downstream task of predicting hyperedges. <br> "Higher-order Bipartite Relations": The third and final part of our thesis is dedicated exclusively to "bipartite hypergraphs": structures that are used to capture higher-order relations between two disjoint node sets, e.g., a patients diagnosis (possibly multiple diseases and symptoms), a movie project (multiple actors and crew members), etc. We first capture the structure of real-world such networks using per-fixed bipartite hypergraphs (those where the set of left and right hyperedges is fixed beforehand), and then focus on the bipartite hyperlink prediction problem. Since existing self-attention based approaches meant for usual hypergraphs do not work for bipartite hypergraphs -- a fact that our experiments validate, we propose a cross-attention model for the same, and use the notion of set-matching over collections of sets to solve for bipartite hyperlink prediction. As a result, we develop a neural network architecture called CATSETMAT that performs way better than any of the other approaches meant to solve the bipartite hyperlink prediction problem. Last but not least, we also explain how, owing to an observation we call the positive-negative dilemma, existing state-of-the-art algorithms fail on bipartite hypergraphs. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjBiZjg4YTgtMWY5Ny00MDNjLWE4MmItMTI0MWRlYzc3MjIz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227cc8bd44-135e-4281-b033-80d7f9df42fd%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjBiZjg4YTgtMWY5Ny00MDNjLWE4MmItMTI0MWRlYzc3MjIz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227cc8bd44-135e-4281-b033-80d7f9df42fd%22%7d</a> DTSTART:20201208T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210115T120000Z UID:f8a5d60ac386b00e944338ba98e88308-111 DTSTAMP:19700101T120015Z DESCRIPTION:On the Round Complexity Landscape of Secure Multi-party Computation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/111/on-the-round-complexity-landscape-of-secure-multi-party-computation/ SUMMARY:In secure multi-party computation (MPC), n parties wish to jointly perform a computation on their private inputs in a secure way, so that no adversary corrupting a subset of the parties can learn more information than their outputs (privacy), nor can they affect the outputs of the computation other than by choosing their own inputs (correctness). The round complexity of MPC protocols is a fundamental question in the area of secure computation and its study constitutes a phenomenal body of work in the MPC literature. The research goal of this thesis is to advance the state of the art by expanding this study of round complexity to various adversarial settings and network models. Following are the main contributions of this thesis organized into three broad categories: <br> MPC for small population ([1,2]). We begin with the study of round-optimal (more generally, round-efficient) MPC protocols for small population, namely involving 3 (3PC) and 4 (4PC) parties tolerating single active corruption (honest majority). On the theoretical side, we settle the exact round complexity of 3PC in honest-majority setting, for a range of security notions such as selective abort, unanimous abort, fairness and guaranteed output delivery. On the practical side, we present efficient, constant-round 3PC and 4PC protocols with fairness and guaranteed output delivery; suitable for high-latency networks such as the Internet. <br> Beyond Traditional Adversaries ([3,4]). We extend the study of round complexity beyond the traditional adversarial settings. First, we overcome the demarcation of study of round complexity of MPC based on resilience (i.e honest majority or dishonest majority settings) and investigate an interesting class of protocols called the Best-of-both-Worlds (BoBW) MPC which simultaneously achieve fairness / guaranteed output delivery in honest majority and unanimous abort in dishonest majority. We nearly settle the question of exact round complexity of BoBW protocols for several popular setups of MPC such as the plain model, Common reference String model (CRS) and PKI model. <br> Second, we overcome the demarcation of study of round complexity of MPC based on single type of corruption i.e either purely passive or active. We consider a generalized adversarial setting where the adversary can simultaneously perform both kinds of corruptions. We settle the question of exact round complexity of MPC protocols achieving fairness and guaranteed output delivery against two such generalized and powerful adversaries called the dynamic and boundary adversaries; in the CRS model. <br> Power of Hybrid Networks ([5]). A popular categorization of study of MPC based on network is the synchronous and asynchronous setting. On one hand, asynchronous networks are more realistic but on the other, synchronous protocols are known to have better fault tolerance and properties compared to their asynchronous counterparts. With the goal of combining their best features, we explore hybrid networks that is asynchronous in nature and yet supports a few synchronous rounds at the onset of a protocol execution. We address fundamental questions that throw light on the minimal synchrony assumption needed to achieve the properties of the fully synchronous protocols. We bridge the existing theoretical feasibility gap between perfectly-secure synchronous and asynchronous VSS and MPC protocols; where verifiable secret sharing (VSS) constitutes a fundamental building block of MPC. <br> References <br> [1] Arpita Patra and Divya Ravi. On the exact round complexity of secure three-party computation. In CRYPTO, 2018. <br> [2] Megha Byali, Arun Joseph, Arpita Patra, and Divya Ravi. Fast secure computation for small population over the internet. In ACM Conference of Computer and Communications Security (CCS), 2018. <br> [3] Arpita Patra, Divya Ravi and Swati Singla. On the Exact Round Complexity of Best-of-both-Worlds Multi-party Computation. In ASIACRYPT, 2020. <br> [4] Arpita Patra and Divya Ravi. Beyond honest majority: The round complexity of fair and robust multi-party computation. In ASIACRYPT, 2019. <br> [5] Arpita Patra and Divya Ravi. On the power of hybrid networks in multi-party computation. IEEE Trans. Inf. Theory, 64(6):4207–4227, 2018. <br> <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NmY0ZGQzOTQtNWJhNS00ZjEyLTg1ODktMmRhOWY1NGNhM2Q1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22a787cc01-57cc-4fc1-b7e1-4e9d51923f6d%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NmY0ZGQzOTQtNWJhNS00ZjEyLTg1ODktMmRhOWY1NGNhM2Q1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22a787cc01-57cc-4fc1-b7e1-4e9d51923f6d%22%7d</a> DTSTART:20210115T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210111T120000Z UID:6470f8e2ae0677626eb553b7b4540d43-112 DTSTAMP:19700101T120008Z DESCRIPTION:Data Science at Scale: Scaling Up by Scaling Down and Out (to Disk) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/112/data-science-at-scale-scaling-up-by-scaling-down-and-out-to-disk/ SUMMARY:The standard solution to scaling applications to massive data is scale-out, i.e., use more computers or RAM. This talk presents my work on complementary techniques: scaling down, i.e., shrinking data to fit in RAM, and scaling to disk, i.e., organizing data on disk so that the application can still run fast. I will describe new compact and I/O-efficient data structures and their applications in stream processing, computational biology, and storage. <br> Concretely, I show how to bridge the gap between the worlds of external memory and stream processing to perform scalable and precise real-time event-detection on massive streams. I show how to shrink genomic and transcriptomic indexes by a factor of two while accelerating queries by an order of magnitude compared to the state-of-the-art tools. I show how to improve file-system random-write performance by an order of magnitude without sacrificing sequential read/write performance. <br> Teams Meeting Link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmExNmNmZjMtODM1Zi00MDUxLWFkNmEtNjdmYThkZWIxNjkx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmExNmNmZjMtODM1Zi00MDUxLWFkNmEtNjdmYThkZWIxNjkx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20210111T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210115T120000Z UID:8bfa3c0efcd823929cb4348b7b5722cf-113 DTSTAMP:19700101T120011Z DESCRIPTION:Discourse-guided Approaches for Event Coreference Resolution URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/113/discourse-guided-approaches-for-event-coreference-resolution/ SUMMARY:Event coreference resolution aims to determine and cluster event mentions that refer to the same real-world event. A typical event coreference resolution system relies on scoring similarity between event-arguments features of all event-pairs in a document followed by clustering. However, a text document follows the principle of language economy where most of the events are mentioned only once (singletons). Consequently, only certain key events that connect other peripheral events are repeated to organize the content and produce a coherent story. Additionally, events are barely mentioned together with all of their arguments which results in a sparse and scattered distribution of coreferential event mentions. In the talk, I will discuss some of my recent works that address sparsity in coreferential event distribution by developing a holistic approach based on the cues from the document content organization structures in news articles. <br> Talk Meeting Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWVjMWI2YTAtNGI1Yy00NGM3LTlhZTEtZWZiMjEzMTY1Mjlj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWVjMWI2YTAtNGI1Yy00NGM3LTlhZTEtZWZiMjEzMTY1Mjlj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20210115T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210121T120000Z UID:f61f444696b6fd1113e03fd30154a7fc-116 DTSTAMP:19700101T120015Z DESCRIPTION:Dismantling the deep neural network black box URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/116/dismantling-the-deep-neural-network-black-box/ SUMMARY:Deep neural networks (DNNs) have been quite successful in a variety of supervised learning tasks. A key reason attributed to the success of DNNs is their ability to automatically learn high level representation of the data. The standard view is that low level representations are learnt in the initial layers, and as one proceeds in depth, sophisticated high level representations are learnt in the deeper layers. In this talk, we will focus on DNNs with rectified linear unit (ReLU) activations (ReLU-DNNs), a widely used sub-class of DNNs. We will exploit the gating property of ReLU activations to build an alternative theory for representation learning in ReLU-DNNs. The highlights are: <br> 1) We encode gating information in a novel neural path feature. We analytically show that the standalone role of gates is characterised by the associated neural path kernel (NPK). <br> 2) We show via experiments (on standard datasets) that almost all useful information is stored in the gates, and that neural path features are learnt during training. <br> 3) We show that the neural path kernel has a composite structure. In case of fully connected DNNs, the NPK is a product of the base kernel, in the case of residual networks with skip connections, the NPK has sum of product (of base kernels) form, and in the case of convolutional nets, the NPK is rotationally invariant. DTSTART:20210121T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210118T120000Z UID:fad933ffee803b21caf489fe88140645-118 DTSTAMP:19700101T120020Z DESCRIPTION:A Theoretical Computer Science Perspective on Consciousness URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/118/a-theoretical-computer-science-perspective-on-consciousness/ SUMMARY:The quest to understand consciousness, once the purview of philosophers and theologians, is now actively pursued by scientists of many stripes. This talk looks at consciousness from the perspective of theoretical computer science. It formalizes the Global Workspace Theory (GWT) originated by cognitive neuroscientist Bernard Baars and further developed by him, Stanislas Dehaene, and others. Our major contribution lies in the precise formal definition of a Conscious Turing Machine (CTM), also called a Conscious AI. We define the CTM in the spirit of Alan Turings simple yet powerful definition of a computer, the Turing Machine (TM). We are not looking for a complex model of the brain nor of cognition but for a simple model of (the admittedly complex concept of) consciousness. After formally defining CTM, we give a formal definition of consciousness in CTM. We then suggest why the CTM has the feeling of consciousness. The reasonableness of the definitions and explanations can be judged by how well they agree with commonly accepted intuitive concepts of human consciousness, the range of related concepts that the model explains easily and naturally, and the extent of its agreement with scientific evidence. Joint work with Manuel Blum and Avrim Blum. DTSTART:20210118T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210121T120000Z UID:a7af7f1fbda2461c6d71f0de44cf57ee-119 DTSTAMP:19700101T120020Z DESCRIPTION:Kicking Butt in Computer Science: Women in Computing at Carnegie Mellon and Around the World URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/119/kicking-butt-in-computer-science-women-in-computing-at-carnegie-mellon-and-around-the-world/ SUMMARY:In this talk we discuss our work on the participation of women in computer science in Carnegie Mellons School of Computer Science. Carnegie Mellon (CMU) has been a national leader in paying attention to womens participation in computing. In the last few years CMU hit a landmark in reaching gender parity in the computer science major. Our book -- “Kicking butt in Computing in Computer Science: Women in Computing at Carnegie Mellon University” -- tells a positive story and illustrates the value of looking to cultural factors, not gender differences. More recently our work moved beyond CMU to ask what is happening with women in computing globally? In “Cracking the Digital Ceiling: Women in Computing Around the World” we brought together more than 20 academics to provide their perspectives on the issue from a wide range of countries and cultures. We will discuss the various obstacles and catalysts that help determine womens participation in the ever growing and life influencing fields of computing. DTSTART:20210121T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210125T120000Z UID:31a830446d238fad914a3c92a68fa5d6-120 DTSTAMP:19700101T120011Z DESCRIPTION:Deep Learning over Hypergraphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/120/deep-learning-over-hypergraphs/ SUMMARY:Though graphs have been extensively used for modelling real-world relational datasets, they are restricted to pairwise relationships, i.e., each edge connects exactly two vertices. Many real-world relational datasets such as academic networks, chemical reaction networks, email communication networks contain group-wise relationships that go beyond pairwise associations. Hypergraphs can flexibly model such datasets by relaxing the notion of an edge to connect an arbitrary number of vertices and providing a mathematical foundation for understanding and learning from large amounts of real-world heterogeneous data. <br> The state-of-the-art techniques for learning from graph data with pairwise relationships use graph-based deep learning models such as graph neural networks. A prominent observation that inspires this thesis is that deep neural networks are still under-explored for hypergraph data with group-wise relationships. Hypergraphs have been utilised as primary data structures in many machine learning tasks such as vertex classification, hypergraph link prediction, and knowledge base completion. However, the fundamental limitation of most existing non-neural techniques is that they cannot leverage high-dimensional features on vertices, especially those which are not present in relational data (e.g., text attributes of documents in academic networks). In this thesis, we propose novel deep learning-based methods for hypergraph data with high dimensional vertex features. <br> 1) Deep Learning for Hypergraph Vertex-level Predictions In the first part of the thesis, we focus on addressing limitations of existing methods for vertex-level tasks over hypergraphs. In particular, we propose HyperGraph Convolutional Network (HyperGCN) for semi-supervised vertex classification over hypergraphs. Unlike existing methods, HyperGCN principally bridges tools from graph neural networks and spectral hypergraph theory. <br> 2) Deep Learning for Hypergraph Link Prediction In the second part, we focus on the task of predicting groupwise relationships (i.e., link prediction over hypergraphs). We propose Neural Hyperlink Predictor (NHP), a novel neural network-based method for link prediction over hypergraphs. NHP uses a novel scoring layer that principally enables us to predict group relationships on incomplete hypergraphs where hyperedges need not represent similarity. <br> 3) Deep Learning for Multi-Relational and Recursive Hypergraphs In the third and final part, we explore more complex structures such as multi-relational hypergraphs in which each hyperedge is typed (i.e., belongs to a relation type) and recursive hypergraphs in which hyperedges can act as vertices in other hyperedges. We first propose Generalised Message Passing Neural Network (G-MPNN) for learning vertex representations on multi-relational ordered hypergraphs. G-MPNN generalises existing MPNNs on graphs, hypergraphs, multi-relational graphs, heterogeneous graphs, and multi-layer networks. We then propose MPNN-Recursive, a novel framework, to handle recursively structured data. Extensive experimentation on real-world hypergraphs show the effectiveness of our proposed models. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MTg4MDZiODQtMWE0Zi00YTA1LWEzMmMtNzU3MzY3Y2E1Zjk5%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2228f0419c-e018-464c-85c0-3f1bd2ac7281%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MTg4MDZiODQtMWE0Zi00YTA1LWEzMmMtNzU3MzY3Y2E1Zjk5%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2228f0419c-e018-464c-85c0-3f1bd2ac7281%22%7d</a> DTSTART:20210125T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210125T120000Z UID:76f7d3118c2316273f416953d4547421-121 DTSTAMP:19700101T120015Z DESCRIPTION:Locally Reconstructable Non-Malleable Secret Sharing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/121/locally-reconstructable-non-malleable-secret-sharing/ SUMMARY:Non-malleable secret sharing (NMSS) schemes, introduced by Goyal and Kumar (STOC 2018), ensure that a secret m can be distributed into shares m1,...,mn (for some n), such that any t (a parameter <= n) shares can be reconstructed to recover the secret m, any t-1 shares doesnt leak information about m and even if the shares that are used for reconstruction are tampered, it is guaranteed that the reconstruction of these tampered shares will either result in the original m or something independent of m. Since their introduction, non-malleable secret sharing schemes sparked a very impressive line of research. In this talk, we present a new feature of local reconstructablility in NMSS, which allows reconstruction of any portion of a secret by reading just a few locations of the shares. This is a useful feature, especially when the secret is long or when the shares are stored in a distributed manner on a communication network. In this talk, we give a compiler that takes in any non-malleable secret sharing scheme and compiles it into a locally reconstructable non-malleable secret sharing scheme. To secret share a message consisting of k blocks of length r each, our scheme would only require reading r + log k bits (in addition to a few more bits, whose quantity is independent of r and k) from each partys share (of a reconstruction set) to locally reconstruct a single block of the message. Microsoft Teans Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTg2OGMwOGQtOTgxYi00OGEyLWE3M2MtOTgzMDNkMGQ0ODUy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%225f3273b8-8838-46b7-b675-b1e9eab4d8ef%22%7d DTSTART:20210125T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210205T120000Z UID:5eb21fc3f47235c90073db2eb2218e5f-122 DTSTAMP:19700101T120021Z DESCRIPTION:Panel Discussion on Gender Inclusivity in Computer Science URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/122/panel-discussion-on-gender-inclusivity-in-computer-science/ SUMMARY: DTSTART:20210205T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210209T120000Z UID:f4654e2af13ba3182b5c6d3c6b4d64ba-123 DTSTAMP:19700101T120012Z DESCRIPTION:Quantum Computing: Current Status and Future Prospects - Part of the IBM Research Distinguished Speaker Series URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/123/quantum-computing-current-status-and-future-prospects-part-of-the-ibm-research-distinguished-speaker-series/ SUMMARY: DTSTART:20210209T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210209T120000Z UID:7ebe8287e37f9119b566fc0ac5b60322-124 DTSTAMP:19700101T120012Z DESCRIPTION:Quantum Computing: Current Status and Future Prospects - Part of the IBM Research Distinguished Speaker Series URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/124/quantum-computing-current-status-and-future-prospects-part-of-the-ibm-research-distinguished-speaker-series/ SUMMARY: DTSTART:20210209T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210215T120000Z UID:2931f71925fab4677aca04e1fe6018fd-125 DTSTAMP:19700101T120012Z DESCRIPTION:Constructing a TLB-based covert channel on GPUs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/125/constructing-a-tlb-based-covert-channel-on-gpus/ SUMMARY:GPUs are now commonly available in most modern computing platforms. They are increasingly being adopted in cloud platforms and data centers due to their immense computing capability. In response to this growth in usage, manufacturers are continuously trying to improve GPU hardware by adding new features. However, this increase in usage and the addition of utility-improving features can create new, unexpected attack channels. In this thesis, we show that two such features—unified virtual memory (UVM) and multi-process service (MPS)—primarily introduced to improve the programmability and efficiency of GPU kernels have an unexpected consequence—that of creating a novel covert timing channel via the GPU’s translation lookaside buffer (TLB) hierarchy. To enable this covert channel, we first perform experiments to understand the characteristics of TLBs present on a GPU. The use of UVM allows fine-grained management of translations, and helps us discover several idiosyncrasies of the TLB hierarchy, such as three-levels of TLB, coalesced entries. We use this newly-acquired understanding to demonstrate a novel covert channel via the shared TLB. We then leverage MPS to increase the bandwidth of this channel by 40×. Finally, we demonstrate the channel’s utility by leaking data from a GPU-accelerated database application. <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3a7f81f3a291db4f6796a0d9cca7ffd68b%40thread.tacv2/1612856627978?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229a1ad18c-768a-4322-8aa7-890013dcb721%22%7d">https://teams.microsoft.com/l/meetup-join/19%3a7f81f3a291db4f6796a0d9cca7ffd68b%40thread.tacv2/1612856627978?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229a1ad18c-768a-4322-8aa7-890013dcb721%22%7d</a> DTSTART:20210215T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210223T120000Z UID:148cd9357bf7b603325e85c5489bdbb6-126 DTSTAMP:19700101T120010Z DESCRIPTION:Online Learning from Relative Subsetwise Preferences URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/126/online-learning-from-relative-subsetwise-preferences/ SUMMARY:The elicitation and aggregation of preferences is often the key to making better decisions. Be it a perfume company wanting to relaunch their 5 most popular fragrances, a movie recommender system trying to rank the most favoured movies, or a pharmaceutical company testing the relative efficacies of a set of drugs, learning from preference feedback is a widely applicable problem to solve. One can model the sequential version of this problem using the classical multiarmed-bandit (MAB) (e.g., Auer, 2002) by representing each decision choice as one bandit-arm, or more appropriately as a Dueling-Bandit (DB) problem (Yue and Joachims, 2009). Although DB is similar to MAB in that it is an online decision making framework, DB is different in that it specifically models learning from pairwise preferences. In practice, it is often much easier to elicit information, especially when humans are in the loop, through relative preferences: `Item A is better than item B is easier to elicit than its absolute counterpart: `Item A is worth 7 and B is worth 4. <br> <br> However, instead of pairwise preferences, a more general subset-wise preference model is more relevant in various practical scenarios, e.g. recommender systems, search engines, crowd-sourcing, e-learning platforms, design of surveys, ranking in multiplayer games. Subset-wise preference elicitation is not only more budget friendly, but also flexible in conveying several types of feedback. For example, with subset-wise preferences, the learner could elicit the best item, a partial preference of the top 5 items, or even an entire rank ordering of a subset of items, whereas all these boil down to the same feedback over pairs (subsets of size 2). The problem of how to learn adaptively with subset-wise preferences, however, remains largely unexplored; this is primarily due to the computational burden of maintaining a combinatorially large, O(n^k), size of preference information in general. <br> <br> We take a step in the above direction by proposing "Battling Bandits (BB)"---a new online learning framework to learn a set of optimal (good) items by sequentially, and adaptively, querying subsets of items of size up to k (k>=2). The preference feedback from a subset is assumed to arise from an underlying parametric discrete choice model, such as the well-known Plackett-Luce model, or more generally any random utility (RUM) based model. It is this structure that we leverage to design efficient algorithms for various problems of interest, e.g. identifying the best item, set of top-k items, full ranking etc., for both in PAC and regret minimization setting. We propose computationally efficient and (near-) optimal algorithms for above objectives along with matching lower bound guarantees. Interestingly this leads us to finding answers to some basic questions about the value of subset-wise preferences: Does playing a general k-set really help in faster information aggregation, i.e. is there a tradeoff between subsetsize-k vs the learning rate? Under what type of feedback models? How do the performance limits (performance lower bounds) vary over different combinations of feedback and choice models? And above all, what more can we achieve through BB where DB fails? <br> <br> We proceed to analyse the BB problem in the contextual scenario – this is relevant in settings where items have known attributes, and allows for potentially infinite decision spaces. This is more general and of practical interest than the finite-arm case, but, naturally, on the other hand more challenging. Moreover, none of the existing online learning algorithms extend straightforwardly to the continuous case, even for the most simple Dueling Bandit setup (i.e. when k=2). Towards this, we formulate the problem of "Contextual Battling Bandits (C-BB)" under utility based subsetwise-preference feedback, and design provably optimal algorithms for the regret minimization problem. Our regret bounds are also accompanied by matching lower bound guarantees showing optimality of our proposed methods. All our theoretical guarantees are corroborated with empirical evaluations. <br> <br> Lastly, it goes without saying, that there are still many open threads to explore based on BB. These include studying different choice-feedback model combinations, performance objectives, or even extending BB to other useful frameworks like assortment selection, revenue maximization, budget-constrained bandits etc. Towards the end we will also discuss some interesting combinations of the BB framework with other, well-known, problems, e.g. Sleeping / Rotting Bandits, Preference based Reinforcement Learning, Learning on Graphs, Preferential Bandit-Convex-Optimization etc. <br> Microsoft Teams link:<br> <a href="https://tinyurl.com/zo46ntdz">https://tinyurl.com/zo46ntdz</a> DTSTART:20210223T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210223T120000Z UID:cb29b5e593c926ea1113f6e579bb5d85-127 DTSTAMP:19700101T120014Z DESCRIPTION:Modern Consensus Protocols: The Synchronous, the Asynchronous, and the Partially Synchronous URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/127/modern-consensus-protocols-the-synchronous-the-asynchronous-and-the-partially-synchronous/ SUMMARY: DTSTART:20210223T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210226T120000Z UID:9eb5fa68b1808a161257d9cb9086ef9f-128 DTSTAMP:19700101T120009Z DESCRIPTION:1. Making Synchronous Consensus Protocols Practical: A Journey 2.Key Management and Zero Knowledge Credentials for Decentralized Identity Ecosystem3.Technical Deep Dive on Hyperledger Fabric4.Ent URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/128/1-making-synchronous-consensus-protocols-practical-a-journey-2-key-management-and-zero-knowledge-credentials-for-decentralized-identity-ecosystem3-technical-deep-dive-on-hyperledger-fabric4-ent/ SUMMARY:1. Byzantine Fault Tolerant protocols in the synchronous setting have often been considered impractical due to the strong synchrony assumption. On the flip side, synchronous protocols can be used to tolerate up to one-half Byzantine faults. In this talk, I will explain my journey towards improving synchronous protocols, both in theory and practice. <br> 2. Decentralized Identity Foundation (DIF) is a collection of international organizations that focuses on building an open ecosystem for self-owned decentralized identity (DID). Microsoft is an important member of DIF and is working on building protocols, infrastructure and open source libraries for the DID ecosystem. As a part of this effort, I worked on building a self owned cryptographic key management scheme and a zero-knowledge credentials scheme. The key management library has already been open sourced. For this talk, I will spend most of time discussing the key management scheme and the wonderful collaborative effort in it that brought together developers, cryptography and security researchers and standards people. Then I will briefly talk about the zero knowledge credentials project highlighting the new requirements and challenges that the DID ecosystem brings out in using traditional zero-knowledge credential schemes. <br> 3. Blockchain is a shared, replicated, immutable transaction ledger which is maintained by a distributed network of nodes. The transactions in the ledger are grouped into blocks that includes a hash that binds the block to its preceding block, thus creating an immutable chain of blocks. Blockchain networks can be primarily categorized into Permissionless and Permissioned networks. In a Permissionless blockchain all the participants are anonymous and hence do not have trust in each other. The only source of trust is that the state of the blockchain, prior to a certain depth, is immutable. On the other hand Permissioned blockchain operates amongst a set of known and identified participants operating under a governance model, which provides a certain degree of trust. This talk will provide a technical deep dive on Hyperledger Fabric, which is an enterprise grade permissioned distributed ledger framework for developing solutions and applications. Hyperledger Fabric has a highly modular and configurable architecture, enabling innovation, versatility and optimization for a broad range of industry use cases including banking, finance, insurance, health care etc. Hyperledger Fabric is the first distributed ledger platform to support smart contracts authored in general-purpose programming languages such as Java, Go and Node.js, rather than constrained domain-specific languages (DSL). Hyperledger Fabric introduces a new architecture for transactions i.e. execute-order-validate, which addresses the resiliency, flexibility, scalability, performance and confidentiality challenges faced by the order-execute model. Hyperledger fabric takes a unique approach to consensus which enables performance and scalability while preserving privacy. <br> 4. Blockchain technology provides greater transparency and security in carrying out business transactions by maintaining immutable transaction records within a distributed network of mutually untrusting entities. A secure distributed consensus protocol is used for maintaining the ledger and blockchain has a framework for automatically executing smart contracts based on the state of the distributed ledger. Blockchaintechnology has been seen as a very promising technology in supply chain as well as financial services industry. Applications related to product traceability, international trade finance, paperless trade, etc. are the initial ones that have gone into production. This talk will provide an overview of blockchain solutions we have developed for various industries. We will also discuss some of the recent trends and interesting research problems in this space. DTSTART:20210226T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210308T120000Z UID:de9cd16d515f0dbcd7550367e9889ec4-129 DTSTAMP:19700101T120014Z DESCRIPTION:Multi-Domain Coupling in Cyber-Physical Systems Design URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/129/multi-domain-coupling-in-cyber-physical-systems-design/ SUMMARY:In a cyber-physical system (CPS), hardware and software components control a physical process. Hence, there is a strong interaction between the physical dynamics, the control law, and the software algorithms and hardware resources enabling computation, communication, sensing, actuation, and data storage. Such systems have become very common in several industry sectors such as automotive, avionics, healthcare, manufacturing, and energy. The state of practice in the industry has been to follow a separation of concerns for CPS designs. That is, the control law is calculated and the hardware/software is developed in isolated stages without sufficient knowledge of each other. Thus, the implementation might not preserve the performance guarantees obtained during the control design. Often this leads to a long integration, testing, and debugging phase, besides producing inferior systems. In many cases, it is also challenging to offer safety guarantees. <br> Considering that many CPSs are safety-critical and cost-sensitive, e.g., modern cars, this talk will advocate the use of integrated modeling, design, and analysis approaches for CPSs. In essence, I will show how the models, metrics, and methods from different engineering domains can be coupled together in a comprehensive framework for the design of safe and cost-efficient CPSs. I will discuss a hybrid optimization technique that enables the co-design of controllers and their distributed software implementations on a realistic automotive hardware platform (i.e., multiple electronic control units connected by a FlexRay bus). I will further illustrate a toolchain, the first of its kind, that integrates the co-design scheme with industry-strength tools for control design and software development respectively. This toolchain enables the design automation for control software. This talk will conclude with a discussion on some promising research directions for next-generation CPSs that need to be secure, adaptive, and autonomous, besides being safe and cost-efficient. <br><br> Teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NTUxNzgxOWUtZGIyNi00Y2NiLThiZjAtMGIxMTljMmJmZTFj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2282f39501-c5b2-4bfb-87c3-f17ca74c00b6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NTUxNzgxOWUtZGIyNi00Y2NiLThiZjAtMGIxMTljMmJmZTFj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2282f39501-c5b2-4bfb-87c3-f17ca74c00b6%22%7d</a> DTSTART:20210308T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210304T120000Z UID:d69550390debc667537ab58966ec6ce9-130 DTSTAMP:19700101T120015Z DESCRIPTION:1. The need for Inclusive STEM Education - ground realities and collaborative solutions <br> 2. Using Data to Build Better Systems and Services URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/130/1-the-need-for-inclusive-stem-education-ground-realities-and-collaborative-solutions-2-using-data-to-build-better-systems-and-services/ SUMMARY:1. India produces one of the highest number of STEM graduates in the world. India also has the highest population of visually impaired persons. However, out of the millions of visually impaired people, less than 50 students have studied STEM subjects beyond high school due to the non-inclusive and largely inaccessible education system. Due to this, people with visual impairments are deprived from choosing the currently flourishing careers in Science and Computing. In this talk, Vidhya will share her experiences in studying STEM subjects as a visually impaired student and role of technologies in enabling independence both in her education and work. She will also share the various initiatives which she and the team at Vision Empower have undertaken to make STEM subjects accessible to visually impaired children over the past 3 years. <br> 2. Today’s systems and services are large and complex, often supporting millions or even billions of users. Such systems are extremely dynamic as developers continuously commit code and introduce new features, fixes and, consequently, new bugs. Multiple problems crop up in such a dynamic environment, from misconfiguration of essential services, very slow testing and deployment procedures, and extended service disruptions when catastrophic bugs hit deployment. Nevertheless, with the advent of cloud-based services, new opportunities to use machine-learning to alleviate such problems have emerged. Large-scale services generate petabytes of code, test, and usage-related data within just a few days. This data can be potentially harnessed to provide valuable insights to engineers on how to improve service performance, security and reliability. However, cherry-picking important information from such vast amounts of systems-related data proves to be a formidable challenge. Over the last few years, we have been working on leveraging code, test logs and telemetry as data to build several tools that help develop and deploy systems faster while maintaining and even improving system reliability. My talk will first describe the challenges that arise from using machine learning on such systems-related data and metadata. Next I will do a deep-dive on the design of a few tools that we built and are being used by several of Microsoft’s services. DTSTART:20210304T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210304T120000Z UID:b14761c4cf529058633d636e099b83b2-132 DTSTAMP:19700101T120017Z DESCRIPTION:PhD in India: Opportunities and Challenges URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/132/phd-in-india-opportunities-and-challenges/ SUMMARY:In this panel discussion, our panellists, all of whom are leading researchers in their respective fields and have obtained their PhD in India, will discuss their research journey -- starting from being PhD students in India to their current positions as entrepreneurs, researchers and academicians, in India and globally. Link for Panel Discussion: Zoom Webinar: https://zoom.us/j/95626106483 DTSTART:20210304T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210311T120000Z UID:3d0a7d38db56268957d8b8d0b7657852-133 DTSTAMP:19700101T120019Z DESCRIPTION:AI in Finance: Scope and Examples URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/133/ai-in-finance-scope-and-examples/ SUMMARY:AI enables principled representation of knowledge, complex strategy optimization, learning from data, and support to human decision making. I will present examples and discuss the scope of AI in our research in the finance domain. DTSTART:20210311T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210310T120000Z UID:3b4870a581737f3657382b46cf2157e8-134 DTSTAMP:19700101T120016Z DESCRIPTION:Statistical Network Analysis: Community Structure, Fairness Constraints,and Emergent Behavior URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/134/statistical-network-analysis-community-structure-fairness-constraintsand-emergent-behavior/ SUMMARY:Networks or graphs provide mathematical tools for describing and analyzing relational data. They are used in biology to model interactions between proteins, in economics to identify trade alliances among countries, in epidemiology to study the spread of diseases, and in computer science to rank webpages on a search engine, to name a few. Each application domain in this wide assortment encounters networks with diverse properties and imposes various constraints. For example, networks may be dynamic, heterogeneous, or attributed, and an application domain may require a fairness constraint on the communities. However, most existing research is concerned with the simplest type of networks with a fixed set of nodes and edges and focuses on the canonical forms of tasks like community detection and link prediction. This thesis aims at bridging this gap by proposing community detection and link prediction methods to analyze different types of networks from various perspectives. <br> Our first contribution is a spectral algorithm with theoretical guarantees that finds 'fair' clusters. We define a notion of individual fairness in communities using an auxiliary representation graph. Nodes are connected in this graph if they can represent each others interests in various communities. Informally speaking, a node considers a community fair if an adequate number of its representatives belong to that community. The goal is to find communities that are considered fair by all nodes under the representation graph. We show that our proposed fairness criterion (i) generalizes the idea of statistical fairness and (ii) is also applicable in cases where the sensitive node attributes (like gender and race) are not observable but instead manifest themselves as intrinsic or latent features of a social network. We develop a fair spectral clustering algorithm and prove that it is weakly consistent (#mistakes = o(n) with probability 1 - o(1)) under a proposed variant of the stochastic block model. <br> Second, we propose a community-based statistical model for dynamic networks where edges appear and disappear over time. Many networks like social networks, citation networks, contact networks, etc., are dynamic in nature. Our model embeds the nodes and communities in a d-dimensional latent space and specifies a procedure for updating these embeddings over time to model the network's evolution. Given an observed dynamic network, we infer these latent quantities using variational inference and use them for link forecasting and community detection. Unlike existing approaches, our model supports the birth and death of communities. It also allows us to use powerful neural networks during inference. Experiments demonstrate that our model is better at link forecasting and community detection as compared to existing approaches. Moreover, it discovers stable communities, as quantified by the normalized mutual information (NMI) score between communities discovered at successive time steps. This desirable quality is absent in methods that ignore the network dynamics. <br> Third, we propose a statistical model for heterogeneous dynamic networks where the nodes and relations additionally have a type associated with them (e.g., knowledge graphs). Besides the latent node attributes, this model also encodes a set of interaction matrices for each type of relation. These matrices specify the affinity between nodes based on their attribute values and can represent both homophilic (like attracts like) and heterophilic relationships (opposites attract). We develop a scalable neural network-based inference procedure for this model and demonstrate that it outperforms existing state-of-the-art approaches on several homogeneous and heterogeneous dynamic network datasets, particularly the temporal knowledge graphs. <br> Fourth, we develop a model for networks with node covariates to bring explainability to community detection. This model integrates node covariates into a stochastic block model using restricted Boltzmann machines. We subscribe to the view that a community can be explained by identifying the defining covariates of its member nodes. Our model provides the relative importance of various covariates for each community, thereby explaining its decision to group the members. Existing approaches for modeling networks with covariates lack this property, especially the ones that are based on deep neural networks. We also derive an efficient inference procedure that runs in linear time in the number of nodes and edges. Experiments confirm that our model's community detection performance is comparable with recent deep neural network-based approaches. However, it additionally offers the advantage of explainability. <br> The discussion till now views communities as passive structures arising out of interactions between nodes. However, just like existing links in a network determine future links, communities also play a functional role in shaping the behavior of the nodes (for example, preference for a clothing brand). Our final contribution explores this functional view of communities and shows that they affect emergent communication in a networked multi-agent reinforcement learning setting. <br> <br> Meeting Link : <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_Yjk2ZWViYWMtYjRhZi00MTdjLWJjNWYtNjAxMGY2MmU3MjEz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c90a12b8-df95-4e40-88fd-ee979f2b42ba%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_Yjk2ZWViYWMtYjRhZi00MTdjLWJjNWYtNjAxMGY2MmU3MjEz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c90a12b8-df95-4e40-88fd-ee979f2b42ba%22%7d</a> DTSTART:20210310T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210319T120000Z UID:2dd7cc1e0e7836480f7ca4d5fe73f8ae-135 DTSTAMP:19700101T120019Z DESCRIPTION:Software Fault Tolerance via Environmental Diversity URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/135/software-fault-tolerance-via-environmental-diversity/ SUMMARY:Time: 7 PM with online Networking at 6.30 PM<br> Meeting Link: <a href="http://bit.do/vvssarma">http://bit.do/vvssarma</a> <br> Complex systems in different domains contain significant amount of software. Several studies have established that a significant fraction of system outages are due to software faults. Traditional methods of fault avoidance, fault removal based on extensive testing/debugging, and fault tolerance based on design/data diversity are found inadequate to ensure high software dependability. The key challenge then is how to provide highly dependable software. We discuss a viewpoint of fault tolerance of software-based systems to ensure high dependability. We classify software faults into Bohr bugs and Mandel bugs, and identify aging-related bugs as a subtype of the latter. Traditional methods have been designed to deal with Bohr bugs. The next challenge then is to develop mitigation methods for Mandel bugs in general and aging-related bugs in particular. We submit that mitigation methods for Mandel bugs utilize environmental diversity. Retry operation, restart application, failover to an identical replica (hot, warm or cold) and reboot the OS are examples of mitigation techniques that rely on environmental diversity. For software aging related bugs it is also possible to utilize proactive environmental diversity technique known as software rejuvenation. We discuss environmental diversity both from experimental and analytic points of view and cite examples of real systems employing these techniques. DTSTART:20210319T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210322T120000Z UID:679e2c34ee5825bcc16133113178bcb2-136 DTSTAMP:19700101T120014Z DESCRIPTION:Revisiting Statistical Techniques for Cardinality Estimation in RDBMS URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/136/revisiting-statistical-techniques-for-cardinality-estimation-in-rdbms/ SUMMARY:Relational Database Management Systems (RDBMS) constitute the backbone of today's information-rich society, providing a congenial environment for handling enterprise data during its entire life cycle of generation, storage, maintenance, and processing. The Structured Query Language (SQL) is the standard interface to query the information present in RDBMS-based storage. Because of the declarative nature of SQL, the query optimizer inside the database engine needs to come up with an efficient execution plan for a given query. To do so, database query optimizers are critically dependent on accurate row-cardinality estimates of the intermediate results generated on the edges of the execution plan tree. <br> Unfortunately, the histogram and sampling-based techniques commonly used in industrial database engines for cardinality estimation are often woefully inaccurate in practice. As a result, query optimizers produce sub-optimal query execution plans, leading to inflated query response times. This lacuna has motivated a recent slew of papers advocating the use of machine-learning techniques for cardinality estimation. However, these new approaches have their own limitations regarding training overheads, output explainability, incorporating dynamic updates, handling of workload drift, and generalization to unseen queries. <br> In this work, we take a relook at the traditional techniques and investigate whether they can be made to work satisfactorily when augmented with light-weight data structures. Specifically, we present GridSam, which essentially combines histograms and sampling in a potent partnership incorporating both algorithmic and platform innovations. <br> From the algorithmic perspective, GridSam first creates a multi-dimensional grid overlay structure by partitioning the data-space on "critical" attributes (Histogram), and then performs dynamic sampling from query-specific regions of the grid to capture correlations (Sampling). A heuristic-based methodology is used to determine the critical grid dimensions. Further, insights from Index-based Join Sampling (IBJS) technique are leveraged to direct the sampling in multi-table queries. Finally, learned-indexes are incorporated to reduce the index-probing cost for join sampling during the estimation process. <br> From the platform perspective, GridSam leverages the massive parallelism offered by current GPU architectures to provide fast grid setup times. This parallelism is also extended to the run-time estimation process. <br> A detailed performance study on benchmark environments indicates that GridSam computes cardinality estimates with accuracies competitive to contemporary learning-based techniques. Moreover, it does so while achieving orders-of-magnitude reduction in setup time. Further, the estimation time is in the same ballpark as both traditional and learning-based techniques. Finally, a collateral benefit of GridSam’s simple design is that, unlike learned estimators, it is natively amenable to dynamic data environments.<br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/channel/19%3aff37417a25ea41889bb3521f22d917be%40thread.tacv2/General?groupId=cf375c71-892f-441a-ab57-27cbe3049dd4&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476">https://teams.microsoft.com/l/channel/19%3aff37417a25ea41889bb3521f22d917be%40thread.tacv2/General?groupId=cf375c71-892f-441a-ab57-27cbe3049dd4&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20210322T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210324T120000Z UID:d0d32daed18ced740293196efc040d62-137 DTSTAMP:19700101T120016Z DESCRIPTION:Towards Efficient Privacy-Preserving Two-Party k-Means Clustering Protocol URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/137/towards-efficient-privacy-preserving-two-party-k-means-clustering-protocol/ SUMMARY:Two-party data mining is a win-win game if played with a guarantee of data privacy from each other. This guarantee is provided by the use of cryptographic techniques in designing the two-party protocol. The need to obtain collaborative data mining results is growing and so is the need for privacy-preserving data mining protocols. Clustering is one of the data mining techniques and one of the popular clustering algorithms is k-means clustering. We studied the recent work for the secure two-party k-means clustering by Bunn and Ostrovosky and found that the protocol is inefficient for practical purposes. The protocol requires communication rounds which are linear in security parameter for the center initialization step and are quadratic in security parameter for an iterative Lloyds step of the k-means clustering algorithm. The challenge in the secure two-party k-means clustering is the exorbitant communication cost occurring due to the high number of interactions between the parties for performing computations on the data. Our work attempts to resolve this problem of inefficiency in k-means clustering protocol in a two-party setting by proposing some modifications. We have come up with two comparison protocols that are required in the k-means clustering protocol. One of the protocols is to find a minimum of two shared numbers which runs in constant communication rounds. Using this protocol as a building block, another protocol is designed to find a minimum of n shared numbers, which runs in O(n) communication rounds. We have also improved a protocol that selects a random value from a domain oblivious to both parties. Apart from this, the idea to avoid the twoparty integer division altogether is incorporated in the k-means clustering protocol. With these improvements, we propose a two-party k-means clustering protocol for which the initialization step requires communication rounds linear in security parameter and Lloyds step requires communications rounds that are independent of the security parameter. The protocol provides a security guarantee in the semi-honest model except for some minor information leakage. We argue that this leakage in the protocol can be tolerated considering the substantial gain in the communication cost We have verified the gain in performance of the modified protocol by implementing both the k-means clustering protocols and running their instances in the same set-up. Microsoft Teams Link: Chim Sonalis M.Tech. Research Colloquium https://teams.microsoft.com/dl/launcher/launcher.html?url=%2F_%23%2Fl%2Fmeetup-join%2F19%3Abd388c3c1f0943eaa1ba3e022b8cd8be%40thread.tacv2%2F1616059672883%3Fcontext%3D%257b%2522Tid%2522%253a%25226f15cd97-f6a7-41e3-b2c5-ad4193976476%2522%252c%2522Oid%2522%253a%25229be1cf83-a6ec-4e58-a5ca-b440a35ae199%2522%257d%26anon%3Dtrue&type=meetup-join&deeplinkId=6855adfa-e655-49be-af11-fa03d2c596c4&directDl=true&msLaunch=true&enableMobilePage=true&suppressPrompt=true DTSTART:20210324T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210324T120000Z UID:42923af042afcea2b5b5725101e8670f-138 DTSTAMP:19700101T120016Z DESCRIPTION:Algorithms for Challenges to Practical Reinforcement Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/138/algorithms-for-challenges-to-practical-reinforcement-learning/ SUMMARY:Reinforcement learning (RL) in real world applications faces major hurdles - the foremost being safety of the physical system controlled by the learning agent and the varying environment conditions in which the autonomous agent functions. A RL agent learns to control a system by exploring available actions. In some operating states, when the RL agent exercises an exploratory action, the system may enter unsafe operation, which can lead to safety hazards both for the system as well as for humans supervising the system. RL algorithms thus need to respect these safety constraints and must do so with limited available information. Additionally, RL autonomous agents learn optimal decisions in the presence of a stationary environment. However, the stationary assumption on the environment is very restrictive. In many real world problems like traffic signal control, robotic applications, etc., one often encounters situations with non-stationary environments, and in these scenarios, RL algorithms yield sub-optimal decisions. <br> We describe algorithmic solutions to the challenges of safety and non-stationary environmental conditions in RL. In order to handle safety restrictions and facilitate safe exploration during learning, we propose a cross-entropy method based sample efficient learning algorithm. This algorithm is developed based on constrained optimization framework and utilizes limited information for the learning of feasible policies. Also, during the learning iterations, the exploration is guided in a manner that minimizes safety violations. The goal of the second algorithm is to find a good policy for control when the latent model of the environment changes with time. To achieve this, the algorithm leverages a change point detection algorithm to monitor change in the statistics of the environment. The results from this statistical algorithm are used to reset learning of policies and efficiently control an underlying system. <br> In the second part of talk, we describe the application of RL to obstacle avoidance problem in UAV quadrotor. Obstacle avoidance in quadrotor aerial vehicle navigation brings in additional challenges in comparison to ground vehicles. Our proposed method utilizes the relevant temporal information available from the ambient surroundings for this problem and adapts attention based deep Q networks combined with generative adversarial networks for this application. <br> <br> Microsoft Teams Link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjIwYjI3YTEtNzU0OC00MDQxLTk1YjAtMzZiMTkzNDY5ZTgz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%221b0586c2-1488-4f8f-ab3c-d2e61940254c%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjIwYjI3YTEtNzU0OC00MDQxLTk1YjAtMzZiMTkzNDY5ZTgz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%221b0586c2-1488-4f8f-ab3c-d2e61940254c%22%7d</a> DTSTART:20210324T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210325T120000Z UID:075428067a324e51d2eb419bd37a8782-139 DTSTAMP:19700101T120014Z DESCRIPTION:nuKSM: NUMA-aware Memory De-duplication for Multi-socket Servers URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/139/nuksm-numa-aware-memory-de-duplication-for-multi-socket-servers/ SUMMARY:Memory management is one of the most critical pieces in an operating system's design. <br> It has several responsibilities ranging from ensuring quick access to data by applications to enabling memory consolidation. For example, judicious placement of pages in multi-socket NUMA (non-uniform memory access) servers could determine the access latencies experienced by an application. Similarly, memory de-duplication can play a pivotal role in memory consolidation and over-commitment. <br> Different responsibilities of memory management can conflict with each other. This often happens when different subsystems of an OS are responsible for different memory management goals, and each works in its silo.<br> In this work, we report one such conflict that appears between memory de-duplication and NUMA management. Linux's memory de-duplication subsystem, namely KSM, is NUMA unaware. We demonstrate that memory de-duplication can have unintended consequences to NUMA overheads experienced by applications running on multi-socket servers. Linux's memory de-duplication subsystem, namely KSM, is NUMA unaware.<br> Consequently, while de-duplicating pages across NUMA nodes, it can place de-duplicated pages in a manner that can lead to significant performance variations, unfairness, and subvert process priority. <br> We introduce NUMA-aware KSM, a.k.a., nuKSM, that makes judicious decisions about the placement of de-duplicated pages to reduce the impact of NUMA and unfairness in execution. nuKSM also enables users to avoid priority subversion. Finally, independent of the NUMA effect, we observed that KSM fails to scale well to large memory systems due to its centralized design. We thus extended nuKSM to adopt a de-centralized design to scale to larger memory. DTSTART:20210325T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210325T120000Z UID:733ec3c3933197c73d74090b4cb89f40-140 DTSTAMP:19700101T120015Z DESCRIPTION:GPM - Exploring GPUs with Persistent Memory URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/140/gpm-exploring-gpus-with-persistent-memory/ SUMMARY:Non-volatile memory (NVM) technologies promise to blur the long-held distinction between memory and storage by enabling durability at latencies comparable to DRAM at byte granularity.<br> Persistent Memory (PM) is defined as NVM accessed via load/store instructions at a fine grain. <br> Due to decade-long research into CPU's software and hardware stack for PM, and with the recent commercialization of NVM under the aegis of Intel Optane, PM's promise of revolutionizing computing seems closer to reality than it has ever been before.<br> Unfortunately, while a significant portion of computation today happens on Graphics Processing Units (GPUs), they are deprived of leveraging PM. We find that there exist GPU-accelerated applications that could benefit from fine-grain persistence. <br> Our key goal is to expose byte-grain persistent memory to GPU kernels. For this, we propose a design for GPU with fine-grained access to PM, a.k.a. GPM which combines commercially available GPUs and NVM through software. We find important use-cases to leverage GPM and create a workload suite called GPMBench. GPMBench consists of 11 GPU-accelerated workloads modified to leverage PM. Finally, we demonstrate the benefits of our proposed design, GPM, over conventional methods of persisting from GPU. DTSTART:20210325T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210330T120000Z UID:e3639d57d6dfb2f03cbc36938a1927dc-141 DTSTAMP:19700101T120016Z DESCRIPTION:Approximation Algorithms for Geometric Packing Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/141/approximation-algorithms-for-geometric-packing-problems/ SUMMARY:We study approximation algorithms for the geometric bin packing problem and its variants. In the two-dimensional geometric bin packing problem (2D GBP), we are given n rectangular items and we have to compute an axis-parallel non-overlapping packing of the items into the minimum number of square bins of side length 1. 2D GBP is an important problem in computer science and operations research arising in logistics, resource allocation, and scheduling. <br> We first study an extension of 2D GBP called the generalized multidimensional bin packing problem (GVBP). Here each item i additionally has d nonnegative weights v_1(i), v_2(i), …, v_d(i) associated with it. Our goal is to compute an axis-parallel non-overlapping packing of the items into bins so that for all j ∈ [d], the sum of the jth weight of items in each bin is at most 1. Despite being well studied in practice, surprisingly, approximation algorithms for this problem have rarely been explored. We first obtain two simple algorithms for GVBP having asymptotic approximation ratios (AARs) 6(d+1) and 3(1 + ln(d+1) + ε). We then extend the Round-and-Approx (R&A) framework [Bansal-Khan, SODA 14] to wider classes of algorithms, and show how it can be adapted to GVBP. Using more sophisticated techniques, we obtain algorithms for GVBP having an AAR of 2(1+ln((d+4)/2))+ε, which improves to 2.919+ε for the special case of d=1. <br> Next, we explore approximation algorithms for the d-dimensional geometric bin packing problem (dD GBP). Caprara (MOR 2008) gave a harmonic-based algorithm for dD GBP having an AAR of 1.69104^(d-1). However, their algorithm doesnt allow items to be rotated. This is in contrast to some common applications of dBP, like packing boxes into shipping containers. We give approximation algorithms for dD GBP when items can be orthogonally rotated about all or a subset of axes. We first give a fast and simple harmonic-based algorithm, called fullh_k, having an AAR of 1.69104^d. We next give a more sophisticated harmonic-based algorithm, which we call hgap_k, having an AAR of (1+eps)1.69104^(d-1). This gives an AAR of roughly 2.860 + ε for 3BP with rotations, which improves upon the best-known AAR of 4.5. In addition, we study the multiple-choice bin packing problem that generalizes the rotational case. Here we are given n sets of d-dimensional cuboidal items and we have to choose exactly one item from each set and then pack the chosen items. Our algorithms fullh_k and hgap_k also work for the multiple-choice bin packing problem. We also give fast and simple approximation algorithms for the multiple-choice versions of dD strip packing and dD geometric knapsack. These algorithms have AARs 1.69104^(d-1) and (1-ε)3^(-d), respectively. <br> A rectangle is said to be δ-thin if it has width at most δ or height at most δ. When δ is a small constant (i.e., close to 0), we give an APTAS for 2D GBP when all rectangles are δ-thin. On the other hand, general 2D GBP is APX-hard. This shows that hard instances arise due to items that are large in both dimensions. <br> A packing of rectangles into a bin is said to be guillotine-separable iff we can use a sequence of end-to-end cuts to separate the items from each other. The asymptotic price of guillotinability (APoG) is the maximum value of opt_G(I)/opt(I) for large opt(I), where opt(I) and opt_G(I) are the minimum number of bins and the minimum number of guillotine-separable bins, respectively, needed to pack I. Computing lower and upper bounds on APoG is an important problem, since proving an upper bound smaller than 1.5 would beat the state-of-the-art algorithm for 2D GBP. The best-known upper bound is 1.69104 and the best-known lower bound is 4/3. We analyze this problem for the special case of δ-thin rectangles, where δ is a small constant (i.e., close to 0). We give a roughly 4/3-asymptotic-approximate algorithm for 2D GBP for this case, which proves an upper-bound of roughly 4/3 on APoG for δ-thin rectangles. We also prove a matching lower-bound of 4/3. This shows that hard examples for upper-bounding APoG include items that are large in both dimensions. <br> <br> Mocrosoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWE5NDlkMDgtNzBiNi00YzYyLWJjNzAtM2QxMzZiOTQ1Mzhi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22cd1ddf68-b75e-4337-87f3-ded65154fa20%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWE5NDlkMDgtNzBiNi00YzYyLWJjNzAtM2QxMzZiOTQ1Mzhi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22cd1ddf68-b75e-4337-87f3-ded65154fa20%22%7d</a> DTSTART:20210330T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210407T120000Z UID:e2277b4a4f6fc1f756a3b1126c2b291b-142 DTSTAMP:19700101T120011Z DESCRIPTION:A Multi-Policy Reinforcement Learning Framework for Autonomous Navigation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/142/a-multi-policy-reinforcement-learning-framework-for-autonomous-navigation/ SUMMARY:Reinforcement Learning (RL) is the process of training an agent to take a sequence of actions with the prime objective of maximizing rewards it obtains from an environment. Deep RL is simply using the same approach where a deep neural network parameterizes the policy. Temporal abstraction in RL is learning useful and generalizable skills, which are often necessary for solving complex tasks in various environments of practical interest. One such domain is the longstanding problem of autonomous vehicle navigation. In this work, we focus on learning complex skills in such environments where the agent has to learn a high-level policy by leveraging multiple skills inside an environment that presents various challenges. Multi-policy reinforcement learning algorithms like the Options Critic Framework require an exorbitant amount of time for converging policies. Even when they do, there is a broad tendency for the policy over options to choose a single sub-policy exclusively, thus rendering the other policies moot. In contrast, our approach combines an iterative approach to complement previously learned policies. To conduct the experiments, a custom simulated 3D navigation environment was developed where the agent is a vehicle that has to learn a policy by which it can avoid a collision. This is complicated because, in some scenarios, the agent needs to infer certain abstract meaning from the environment to make sense of it while learning from a reward signal that becomes increasingly sparse. In this thesis, we introduce the `Stay Alive' approach to learn such skills by sequentially adding them into an overall set without using an overarching hierarchical policy where the agent's objective is to prolong the episode for as long as possible. The general idea behind our approach comes from the fact that both animals and human beings learn meaningful skills in previously acquired skills to better adapt to their respective environments. We compare and report our results on the navigation environment and the Atari Riverraid environment with state-of-the-art RL algorithms and show that our approach outperforms the prior methods. Microsoft Meeting Link : https://teams.microsoft.com/l/meetup-join/19%3ameeting_YTI5M2MzOWMtMDEwNS00MzU4LTgyN2MtNWZmNGYzMTk0YjQ0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c90a12b8-df95-4e40-88fd-ee979f2b42ba%22%7d DTSTART:20210407T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210428T120000Z UID:7bc8d0c0c1a917cf68c8eaf3e5ed98ea-145 DTSTAMP:19700101T120010Z DESCRIPTION:A Novel Neural Network Architecture for Sentiment-oriented Aspect-Opinion Pair Extraction URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/145/a-novel-neural-network-architecture-for-sentiment-oriented-aspect-opinion-pair-extraction/ SUMMARY:Over the years, fine-grained opinion mining in online reviews has received great attention from the NLP research community. It involves different tasks such as Aspect Term Extraction (ATE), Opinion Term Extraction (OTE), etc. Opinion Term Extraction (OTE) aims to detect different opinion expressions which convey certain attitude in the review while Aspect Term Extraction (ATE) aims to identify the entities or proposition from the review at which the attitude is directed. Recently, the NLP research community got attracted to aspect-opinion relation modeling. Such modeling would be helpful for aspect-opinion pair extraction that would be used for downstream tasks such as aspect-based sentiment analysis, opinion summarization, etc. <br> As online reviews may contain different sentiment polarities for different aspects of the products, it would help companies find all aspects for which the customers gave positive or negative feedback. In this thesis, we propose a new opinion mining task called Sentiment-oriented Aspect-Opinion Pair Extraction (SAOPE), which aims to find all aspect-opinion pairs from customer reviews given that these pairs convey the specified sentiment polarity. <br> We present a novel neural network architecture for the SAOPE task. In the proposed approach, aspect-opinion co-extraction is performed first and then the aspect-opinion pairs are generated through relation modeling. The aspect and the corresponding opinion words are closely related in the dependency trees. Hence, we explore graph neural networks to utilize syntactic information generated from the dependency tree of the reviews to model the relationship between the aspects and corresponding opinion words. We design a modified graph attention network (GAT) called Graph Co-attention Network (GCAT) and compare its performance with Graph Convolution Network (GCN) and Graph Attention Network (GAT) for the aspect-opinion co-extraction and the relation detection. For the SAOPE task, we evaluate our model on SemEval Challenge datasets and show that GCAT and GAT perform better than the baseline model with GCN for aspect-opinion co-extraction. We demonstrate that the proposed Graph Co-attention Network (GCAT) performs better than other graph neural networks for aspect-opinion relation detection on the publicly available benchmark datasets. <br> Microsoft Teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWMzMGM4NzItYTliNC00NmIzLWIxZDQtNTkzNTU5OGFiZTI3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWMzMGM4NzItYTliNC00NmIzLWIxZDQtNTkzNTU5OGFiZTI3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20210428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210426T120000Z UID:ca47b6791e1eb1896e4a1ac04828f446-146 DTSTAMP:19700101T120016Z DESCRIPTION:Scaling Blockchains Using Coding Theory and Verifiable Computing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/146/scaling-blockchains-using-coding-theory-and-verifiable-computing/ SUMMARY:The issue of scalability has been restricting blockchains from their widespread adoption. The current transaction rate of bitcoin is around seven transactions/second while the blockchain size has crossed the 300 GB mark. Although many different approaches have been proposed to scale blockchains, e.g., sharding, lightning network, etc., we focus our analysis on methods utilizing ideas from coding theory and verifiable computing. We first consider SeF, a blockchain archiving architecture utilizing LT codes to reduce storage constraints per node up to 1000x. SeF enables full nodes to store only a small number of encoded blocks or droplets instead of an entire blockchain. Although efficient in the average case, the architecture sometimes requires large bandwidth (many droplets) to reconstruct blockchain. While other rate-less coding strategies utilizing two encoding levels are proven better than LT codes, we investigate their suitability in the proposed architecture. We propose and simulate three techniques about how to incorporate these coding strategies. The results show that precode-based rate-less coding schemes provide similar storage savings with reduced bandwidth variance for recovery. The other work we examine is PolyShard, which introduces the notion of coded-sharding. Coded sharding exports block verification of sub-ledger to the whole network instead of nodes handling that sub-ledger, making sharding resilient even to an adaptive adversary, i.e., adversary having the power to corrupt nodes after their assignment to shards. However innovative, PolyShard requires decoding of Reed-Solomon codes over large fields for block verification in real-world settings, making it computationally intensive and less practical. We propose replacing the decoding phase with verifiable computing, which reduces the bottleneck and makes the system practical for light verification functions. <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/channel/19%3a90c3a6855afe407d9a0516f42cff8e4c%40thread.tacv2/General?groupId=ae7f9d05-a7e4-423b-8bf4-800e78978105&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476"> https://teams.microsoft.com/l/channel/19%3a90c3a6855afe407d9a0516f42cff8e4c%40thread.tacv2/General?groupId=ae7f9d05-a7e4-423b-8bf4-800e78978105&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476</a> DTSTART:20210426T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210427T120000Z UID:ffe287f21acf671a657cfd5e9d3b9364-147 DTSTAMP:19700101T120009Z DESCRIPTION:Novel First-order Algorithms for Non-smooth Optimization Problems in Machine Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/147/novel-first-order-algorithms-for-non-smooth-optimization-problems-in-machine-learning/ SUMMARY:This thesis is devoted to the design of efficient optimization algorithms for machine learning (ML) problems where the underlying objective function to be optimized is convex but not necessarily differentiable. Such non-smooth objective functions are ubiquitous in ML mainly due to the use of one or more of the following: (a) non-differentiable loss functions, (b) sparsity promoting regularization terms, (c) constraint sets to induce specific structure on the parameters to be learned. Motivated by a wide range of learning problems that can be cast as optimization of a non-smooth convex objective, we focus on developing first-order algorithms with non-asymptotic convergence rate guarantees to solve such problems in large-scale settings. Based on shortcomings of the existing research in this domain, we address the following specific issues in this thesis. <br> First, we consider the problem of learning a kernel matrix from m similarity matrices under a general convex loss. The existing algorithms do not apply if one employs arbitrary loss functions and often can not handle m>1 case. Based on the Mirror Descent (MD) framework, we present several provably convergent iterative algorithms that exploit the availability of off-the-shelf support vector machine (SVM) solvers. One of the significant contributions involves an extension of the well-known MD algorithm for simplex to handle the Cartesian product of positive semidefinite (PSD) matrices leading to a novel algorithm called Entropic Multiple Kernel Learning. We also show simulation results on protein structure classification involving several similarity matrices to demonstrate the proposed algorithms efficacy. <br> Next, we focus on minimizing a convex function over a feasible set given by the intersection of finitely many simple sets, each of which is equipped with a projection oracle. Examples of constraint sets that possess such structure include the set of doubly stochastic matrices, elliptope, the intersection of PSD cone with an L1-norm ball, etc. The main difficulty lies in computing the projection of a point onto the feasible set. Exploiting the intersecting sets linear regularity property, we present an exact penalty approach that leads to first-order algorithms with explicit guarantees on the approximate solutions distance from the feasible set. Further, we show improved iteration-complexity when the objective possesses structural smoothness / strong convexity. This is achieved through a saddle-point reformulation where the proximal operators required by the primal-dual algorithms can be computed in closed form. We illustrate the benefits of our approach on a graph transduction problem and on graph matching. <br> Third, we consider convex-concave saddle point problems with bilinear interaction structure. This class of problems encompasses most convex optimization problems arising in ML and includes minimizing the sum of many simple non-smooth convex functions as a special case; thereby, it subsumes learning problems involving complex regularization terms such as total-variation based image denoising, overlapping group lasso, isotonic regression, etc. We first propose a primal-dual algorithm for this general class of problems that can achieve the O(1/T) convergence rate guarantee on the non-ergodic primal-dual iterate pair. Further, assuming strong convexity in the primal domain, we show an improved non-ergodic convergence rate of O(1/T^2). In contrast, the existing primal-dual algorithms achieve such convergence rate only in the ergodic or semi-ergodic setting. <br> Finally, we consider the classical setting of minimizing the sum of two convex functions: a smooth one (possessing Lipschitz continuous gradient) and a simple non-smooth one with easy to compute proximal operator. The well-known FISTA algorithm (also Nesterovs accelerated gradient method) achieves the optimal O(1/T^2) non-ergodic convergence rate for this problem. One of the drawbacks of these fast gradient methods is that they require computing gradients of the smooth function at points different from those on which the convergence rate guarantee applies. Inspired by the use of past gradients as momentum in training deep nets, we propose an accelerated gradient algorithm to rectify this drawback keeping the convergence rate intact. We achieve this through a judicious choice of momentum in both primal and dual space. To the best of our knowledge, this is the first accelerated gradient algorithm that achieves an O(1/T^2) convergence rate guarantee on the iterates at which gradients are evaluated. This fills a significant research gap as Polyaks Heavy Ball method guarantees accelerated convergence rate through gradient momentum only for a restrictive class of twice differentiable and strongly convex objective functions. <br> The Microsoft Teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NmFmY2Q2ZmYtYTEwZC00YmZkLWE3NDMtNGNkMTdlNzEwY2Fk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22f9260723-90fc-447a-9c09-fe900a1e6645%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_NmFmY2Q2ZmYtYTEwZC00YmZkLWE3NDMtNGNkMTdlNzEwY2Fk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22f9260723-90fc-447a-9c09-fe900a1e6645%22%7d</a> DTSTART:20210427T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210428T120000Z UID:7a2001f285524ab8ff48400c637e5bd7-148 DTSTAMP:19700101T120011Z DESCRIPTION:A Syntactic Neural Model for Question Decomposition URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/148/a-syntactic-neural-model-for-question-decomposition/ SUMMARY:Question decomposition along with single-hop Question Answering (QA) system serve as useful modules in developing multi-hop Question Answering systems, mainly because the resulting QA system is interpretable and has been demonstrated to exhibit better performance. The problem of Question Decomposition can be posed as a machine translation problem and it can be solved using any sequence-to-sequence neural architecture. Using this approach, it is difficult to capture the innate hierarchical structure of the decomposition. Inspired by database query languages a pseudo-formalism for capturing the meaning of questions, called Question Decomposition Meaning Representation (QDMR) was recently introduced. In this approach, a complex question is decomposed into simple queries which are mapped into a small set of formal operations. This method does not utilize the underlying syntax information of QDMR to generate the decomposition. <br> In the area of programming language code generation, methods that use syntax information as a prior knowledge have been demonstrated to perform better. Moreover, the syntax-aware models are usually interpretable. Motivated by the success of syntax-aware models, we propose a new syntactic neural model for question decomposition in this thesis. In particular, we encode the underlying syntax of the QDMR structures into a grammar model as a sequence of actions. This is done using a deterministic framework which uses Abstract Syntax Trees (AST) and Parse Trees. The proposed approach can be thought of as an encoder-decoder method for QDMR structures where a sequence of possible actions is a latent representation of the QDMR structure. The advantage of using this latent representation is that it is interpretable. Experimental results on a real-world dataset demonstrate that the proposed approach outperforms the state-of-the-art approach especially in scenarios where training data is limited. Some heuristics to further improve the performance of the proposed approach are also suggested in this work. <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ODk1OTI2YjEtNTg5NS00NGQ2LWExYjUtZjVkNDc5ODNhOTZm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ODk1OTI2YjEtNTg5NS00NGQ2LWExYjUtZjVkNDc5ODNhOTZm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20210428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210503T120000Z UID:c9adb363d8227cc08367003b3de1ff95-149 DTSTAMP:19700101T120011Z DESCRIPTION:Robust Algorithms for recovering planted structures in Semi-random instances URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/149/robust-algorithms-for-recovering-planted-structures-in-semi-random-instances/ SUMMARY:In this work, we design algorithms for two fundamentally important and classical graph problems in the planted setting. Both of these problems are NP-hard and the bounds known from the algorithmic front are either not fully understood, or not much progress can be made because of tight lower bounds. Thus it is natural to consider semi-random models for these problems. These models are inspired from the seminal paper of Feige and Killian [FK01] and have been studied in numerous follow-up works with the latest ones by Steinhardt, and McKenzie et al. [Ste17, MMT20]. The construction of our instance starts with an empty graph, then an arbitrary set of vertices (S) is chosen and either a dense graph or a clique (or an independent set) is planted on it, the subgraph on S x VS is a random graph, while the subgraph on VS might be a random, arbitrary, or some special graph (depending on the model). Our algorithms are based on rounding semidefinite programs and our primary focus is on recovering (completely or partially) the planted structure (S) with high probability (over the randomness of the input). We give algorithms that exploit the geometry of the corresponding vectors (from the SDP) and are easy to design/analyse. <br> The two problems which we study are: <br> 1) Densest k-Subgraph/k-Clique Given an undirected graph G, the Densest k-Subgraph problem (DkS) asks to compute a set S subseteq V of cardinality k such that the weight of edges inside S is maximized. This is a fundamental NP-hard problem whose approximability, in spite of many decades of research, is yet to be settled. There is a significant gap between the best known worst-case approximation factor of this problem [BCC+10] and the hardness of approximation for it (assuming the Exponential Time Hypothesis) [Man17]. We ask what are some DTSTART:20210503T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210511T120000Z UID:ce87babace50cac37ef3edbd6952f8e1-150 DTSTAMP:19700101T120016Z DESCRIPTION:New Algorithmic and Hardness results in Learning, Error Correcting Codes, and Constraint Satisfaction Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/150/new-algorithmic-and-hardness-results-in-learning-error-correcting-codes-and-constraint-satisfaction-problems/ SUMMARY:Approximation algorithms are a natural way to deal with the intractability barrier that is inherent in many naturally arising computational problems. However, it is often the case that the task of solving the approximation version of the problem is as hard as the exact version of the problem itself, which is the underlying philosophy driving the study of hardness of approximation. While the last couple of decades has seen significant progress in terms of tight inapproximability results for many important problems, there are still several fundamental problems for which best known lower bounds in terms of the hardness of approximation is still quite far off from their best known upper bounds achievable using efficient algorithms. In this thesis, we investigate this phenomenon in the context of several fundamental com- putational problems in Learning Theory, Error Correcting Codes and Constraint Satisfaction Problems (CSPs), along the following themes: <br> 1. Hardness of Learning using Thresholds: Improper learning – i.e., learning a concept class with a different and potentially easier to deal with hypothesis class – is often used as natural relaxation when the exact (proper) learning problem is intractable. In this thesis, we show the NP-Hardness of (improperly) learning natural concept classes such as Halfspaces and Disjunctive Normal Forms using threshold functions. Our results are tight with respect to efficiently achievable approximation factors for learning using arbitrary functions of thresholds. <br> 2. Parameterized Complexity of Finding Sparse Solutions: Many natural problems such as k-VectorSum and k-EvenSet can be modeled as finding a sparse solution to a system of equations over finite fields. We study these problems in the framework of parameterized complexity and give new (and often tight) lower bounds for exact and approximation versions of these problems under various hypotheses. We also use these results to establish new hardness results for learning sparse parities. <br> 3. Vertex Deletion CSPs: We study Vertex deletion CSPs, which are a variant of CSPs where the objective is to delete the least number of vertices to recover a fully satisfi- able sub-instance. In particular, we give new algorithmic and hardness results for the StrongUniqueGames problem which is the vertex deletion variant of UniqueGames, establishing a connection with small-set-vertex-expansion en route. We also give new algo- rithmic results for vertex deletion CSPs such as OddCycleTransversal, Balanced- VertexSeparator, Partial-3-Coloring in more tractable settings such as when the underlying constraint graph has low threshold rank or is sampled semi-randomly. <br> 4. Testing Sparsity: We also investigate the following question: can one design efficient algorithms for testing a property for which the search/optimization version is intractable? We study this question in the form of testing sparsity, where one simply wishes to check whether a matrix admits a sparse factorization. While the optimization of the problem (Dictionary Learning) is intractable without distributional assumptions, we design an efficient algorithm for the property testing formulation of this problem. We also extend our results to the noise tolerant setting and settings where the basis of representation is known. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OGQzZDlkODItZDBmNi00M2JmLWJjYmEtZmRjYWJkOTU4NWQy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22497b6b67-f9b1-41bc-a3fb-f22b0fd477ef%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OGQzZDlkODItZDBmNi00M2JmLWJjYmEtZmRjYWJkOTU4NWQy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22497b6b67-f9b1-41bc-a3fb-f22b0fd477ef%22%7d</a> DTSTART:20210511T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210528T120000Z UID:f37601b9d0d49ac55a0b032c140c9c00-151 DTSTAMP:19700101T120016Z DESCRIPTION:Can Non-Humanoid Social Robots Reduce Workload of Special Educators : An Online and In-Premises Field Study (Joint work with Academy for severe handicaps and Autism) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/151/can-non-humanoid-social-robots-reduce-workload-of-special-educators-an-online-and-in-premises-field-study-joint-work-with-academy-for-severe-handicaps-and-autism/ SUMMARY:Socially Assistive Robotics have been used in Autism Spectrum Disorder (ASD) interventions, such studies often exclude Special Educators (SEs) and often use expensive humanoid robots. In this paper, we investigate whether non-humanoid toy robots can act as teaching aids in ASD Education, in particular, can they reduce the workload of SEs. We target two most common yet divergent problems from Individualized Education Plans (IEPs) of ASD children - communication and gross motor skills. We present results from three studies a) toy robot Cozmo assists SEs in verbal lessons in school premises, b) mini drone Tello helps SEs in exercise lessons in school premises, and c) Cozmo, SEs, and ASD children connect remotely, as mandated due to the Covid-19 pandemic, for verbal lessons. All three studies showed improvement in learning outcomes and reduction in prompts from the SEs, denoting reduced workload. The effect of a robots virtual presence in online ASD interventions has not been studied before. However, our results show that children spent more time on lessons in online intervention with Cozmo, suggesting that using robots should also be considered when designing online interventions. Furthermore, the roles of Cozmo were analyzed, and we found children showed increased spontaneous interaction when Cozmo acts as a Co-Instructor. Thus, preliminary results indicate toy robots, as opposed to expensive humanoids, may have significant potential in aiding SEs in Autism education. DTSTART:20210528T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210604T120000Z UID:fa297dd3550dd29681cfa109ef0a2167-152 DTSTAMP:19700101T120011Z DESCRIPTION:A Trusted-Hardware Backed Secure Payments Platform for Android URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/152/a-trusted-hardware-backed-secure-payments-platform-for-android/ SUMMARY:Digital payments using personal electronic devices have been steadily gaining in popularity for the last few years. While digital payments using smartphones are very convenient, they are also more susceptible to security vulnerabilities. Unlike devices dedicated to the purpose of payments (e.g. POS terminals), modern smartphones provide a large attack surface due to the presence of so many apps for various use cases and a complex feature-rich smartphone OS. Because it is the most popular smartphone OS by a huge margin, Android is the primary target of attackers. Although the security guarantees provided by the Android platform have improved signifi cantly with each new release, we still see new vulnerabilities being reported ever month. Vulnerabilities in the underlying Linux kernel are particularly dangerous because of their severe impact on app security. To protect against a compromised kernel, some critical functions of the Android platform such as cryptography and local user authentication have been moved to a Trusted Execution Environment (TEE) in the last few releases. But the Android platform doesn't yet provide a way to protect a user's con fidential input meant for a remote server, from a compromised kernel. Our work aims to address this gap in Android's use of TEEs for app security. We have designed a framework that leverages a TEE for protecting user's confi dential input and we have shown how this framework can be used to improve the security of digital payments. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OGU1ZmRjNWUtM2Q2ZC00NzcwLTlhZWQtMDJhMDlkYWIzNzcy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OGU1ZmRjNWUtM2Q2ZC00NzcwLTlhZWQtMDJhMDlkYWIzNzcy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d</a> DTSTART:20210604T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210609T120000Z UID:e34d4e268099b54fcb689c2b8c93e249-153 DTSTAMP:19700101T120011Z DESCRIPTION:Design, Implementation, and Analysis of a TLB-based Covert Channel on GPUs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/153/design-implementation-and-analysis-of-a-tlb-based-covert-channel-on-gpus/ SUMMARY:GPUs are now commonly available in most modern computing platforms. They are increasingly being adopted in cloud platforms and data centers due to their immense computing capability. In response to this growth in usage, manufacturers are continuously trying to improve GPU hardware by adding new features. However, this increase in usage and the addition of utility-improving features can create new, unexpected attack channels. In this thesis, we show that two such features—unified virtual memory (UVM) and multi-process service (MPS)—primarily introduced to improve the programmability and efficiency of GPU kernels have an unexpected consequence—that of creating a novel covert timing channel via the GPUs translation lookaside buffer (TLB) hierarchy. To enable this covert channel, we first perform experiments to understand the characteristics of TLBs present on a GPU. The use of UVM allows fine-grained management of translations, and helps us discover several idiosyncrasies of the TLB hierarchy, such as three-levels of TLB, coalesced entries. We use this newly-acquired understanding to demonstrate a novel covert channel via the shared TLB. We then leverage MPS to increase the bandwidth of this channel by 40×. Finally, we demonstrate the channels utility by leaking data from a GPU-accelerated database application. <br> <br> Microsoft Teams - ONLINE <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjJhMWJjYTYtNGRlZi00ODFmLWI3OTQtZTIyOGVmZDc5NzA2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjJhMWJjYTYtNGRlZi00ODFmLWI3OTQtZTIyOGVmZDc5NzA2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d</a> DTSTART:20210609T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210610T120000Z UID:3307633564dea2b0c0b5aee8d3658d38-154 DTSTAMP:19700101T120012Z DESCRIPTION:Revisiting Statistical Techniques for Result Cardinality Estimation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/154/revisiting-statistical-techniques-for-result-cardinality-estimation/ SUMMARY:The Relational Database Management Systems (RDBMS) constitute the backbone of today's information-rich society, providing a congenial environment for handling enterprise data during its entire life cycle of generation, storage, maintenance, and processing. The Structured Query Language (SQL) is the standard interface to query the information present in RDBMS-based storage. Knowing the expected size of the SQL query result, measured in terms of the output row-cardinality, prior to execution can benefit both the RDBMS system and the user in several ways. The use-cases include assessing query feasibility, approximate query answering, query progress monitoring, and resource allocation strategies. In the context of our work, we define cardinality estimation as the estimation of the result size (number of rows in the output) of the given SQL query. Unfortunately, the histogram and sampling-based techniques commonly used in industrial database engines for cardinality estimation are often woefully inaccurate in practice. This lacuna has motivated a recent slew of papers advocating the use of machine-learning/deep-learning techniques for cardinality estimation. However, these new approaches have their own limitations regarding significant training effort, inability to handle dynamic data-updates, and generalization to unseen queries. In this work, we take a relook at the traditional random sampling and investigate whether they can be made to work satisfactorily when augmented with lightweight data structures. Specifically, we present GridSam – a Grid-based Dynamic Sampling technique, which essentially augments random sampling with histograms, incorporating both algorithmic and platform innovations. From the algorithmic perspective, GridSam first creates a multi-dimensional grid overlay by partitioning the data-space on “critical” attributes, and then performs dynamic sampling from the confined query-specific region of the grid to capture correlations. A greedy methodology targeted towards reducing the Zero Sample Problem occurrence is used to determine the set of “critical” attributes as the grid dimensions. Further, insights from the Index-based Join Sampling (IBJS) technique are leveraged to direct the sampling in multi-table queries. Finally, learned-indexes are incorporated to reduce the index-probing cost for join sampling during the estimation process. From the platform perspective, GridSam leverages the massive parallelism offered by current GPU architectures to provide fast grid setup times. This parallelism is also extended to the run-time estimation process. A detailed performance study on benchmark environments indicates that GridSam computes cardinality estimates with accuracies competitive to contemporary learning-based techniques. Moreover, it does so while achieving an orders-of-magnitude reduction in setup time. Further, the estimation time is in the same ballpark as both traditional and learning-based techniques. Finally, a collateral benefit of GridSam's simple and highly parallelizable design is that, unlike learned estimators, it is amenable to dynamic data environments with frequent data-updates. Microsoft Teams Link: https://teams.microsoft.com/l/channel/19%3ahnmvZ5nGAP1ZF26uEVJEq8xHNmslkLLM1laDUK5iHVk1%40thread.tacv2/General?groupId=b3316384-7d8e-441d-80c8-fd1d1a5250c6&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476 DTSTART:20210610T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210604T120000Z UID:d123558cf7fe031c5c1d9d258f018c3b-155 DTSTAMP:19700101T120016Z DESCRIPTION:Analysis and Methods for Knowledge Graph Embeddings URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/155/analysis-and-methods-for-knowledge-graph-embeddings/ SUMMARY:Knowledge Graphs (KGs) are multi-relational graphs where nodes represent entities, and typed edges represent relationships among entities. These graphs store real-world facts such as (Lionel Messi, plays-for-team, Barcelona) as edges, called triples. KGs such as NELL, YAGO, Freebase, and WikiData have been very popular and support many AI applications such as Web Search, Query Recommendation, and Question Answering. Although popular, these KGs suffer from incompleteness. Learning Knowledge Graph Embeddings (KGE) is a popular approach for predicting missing edges (i.e., link prediction) and representing entities and relations in downstream tasks. While numerous KGE methods have been proposed in the past decade, our understanding and analysis of such embeddings have been limited. Further, such methods only work well with ontological KGs. In this thesis, we address these gaps. <br> Firstly, we study various KGE methods and present an extensive analysis of these methods, resulting in interesting insights. Next, we address an under-explored problem of link prediction in Open Knowledge Graphs (OpenKGs) and present a novel approach that improves the type compatibility of predicted edges. Lastly, we present an adaptive interaction framework for learning KG embeddings that generalizes many existing methods. <br> <br> Analysis of KGE Embeddings In the first part, we present a macro and a micro analysis of embeddings learned by various KGE methods. <br> Despite the popularity and effectiveness of KG embeddings, their geometric understanding (i.e., arrangement of entity and relation vectors in vector space) is unexplored. We initiate a study to analyze the geometry of KG embeddings and correlate it with task performance and other hyper-parameters. Firstly, we present a set of metrics (e.g., Conicity, ATM) to analyze the geometry of a group of vectors. Using these metrics, we find sharp differences between the geometry of embeddings learned by different classes of KGE methods. The vectors learned by a multiplicative model lie in a narrow cone, unlike additive models where the vectors are spread out in the space. This behavior of multiplicative models is amplified by increasing the number of negative samples used for training. Further, a very high Conicity value is negatively correlated with the performance on the link prediction task. <br> We also study the problem of understanding KG embeddings semantics and propose an approach to learn more coherent dimensions. A dimension is coherent if the top entities have similar types (e.g., person). In this work, we formalize the notion of coherence using entity co-occurrence statistics and propose a regularizer term that maximizes coherence while learning KG embeddings. The proposed approach significantly improves coherence while having a comparable performance with baseline in the link prediction and triple classification tasks. Further, based on the human evaluation, we demonstrate that the proposed approach learns more coherent dimensions than the baseline. <br> <br> A method for OpenKG Embedding In the second part, we address the problem of learning KG embeddings for Open Knowledge Graphs (OpenKGs), focusing on improving link prediction. An OpenKG refers to a set of (head noun phrase, relation phrase, tail noun phrase) triples such as (tesla, return to, new york) extracted from a text corpus using OpenIE tools. While OpenKGs are easy to bootstrap for a domain, they are very sparse. Therefore, link prediction becomes an important step while using these graphs in downstream tasks. Learning OpenKG embeddings is one approach for link prediction that has received some attention lately. However, on careful examination, we find that current algorithms often predict noun phrases (NPs) with incompatible types for given noun and relation phrases. We address this problem and propose OKGIT that improves OpenKG link prediction using novel type compatibility score and type regularization. With extensive experiments on multiple datasets, we show that the proposed method achieves state-of-the-art performance while producing type compatible NPs in the link prediction task. <br> <br> An Adaptive Framework for KG Embeddings In the third part, we address the problem of improving the KGE models. <br> Firstly, we show that the performance of existing approaches varies across different datasets, and a simple neural network-based method can consistently achieve better performance on these datasets. Upon analysis, we find that KGE models depend on fixed sets of interactions among the dimensions of entity and relation vectors. Therefore, we investigate ways to learn such interactions automatically during training. We propose an adaptive interaction framework for learning KG embeddings, which can learn appropriate interactions while training. We show that some of the existing models could be seen as special cases of the proposed framework. Based on this framework, we also present two new models, which outperform the baseline models on the link prediction task. Further analysis demonstrates that the proposed approach can adapt to different datasets by learning appropriate interactions. <br> <br> Microsoft Teams Link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MDdmNzk0ZTMtZGRjOS00ZmExLTgwZjQtZmI1YmNlZDk3OWVm%40thread.v2/0?context=%7b">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MDdmNzk0ZTMtZGRjOS00ZmExLTgwZjQtZmI1YmNlZDk3OWVm%40thread.v2/0?context=%7b</a> DTSTART:20210604T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210616T120000Z UID:ee957e8b5a0cc8aba1fc829393df24f4-157 DTSTAMP:19700101T120011Z DESCRIPTION:MPCLeague: Robust MPC Platform for Privacy-Preserving Machine Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/157/mpcleague-robust-mpc-platform-for-privacy-preserving-machine-learning/ SUMMARY:In the modern era of computing, machine learning tools have demonstrated their potential in vital sectors, such as healthcare and finance, to derive proper inferences. The sensitive and confidential nature of the data in such sectors raises genuine concerns for data privacy. This motivated the area of Privacy-preserving Machine Learning (PPML), where privacy of data is guaranteed. Typically, machine learning techniques require significant computing power, which leads clients with limited infrastructure to rely on the method of Secure Outsourced Computation (SOC). In the SOC setting, the computation is outsourced to a set of specialized and powerful cloud servers and the service is availed on a pay-per-use basis. In this thesis, we design an efficient platform, MPCLeague, for PPML in the SOC setting using Secure Multi-party Computation (MPC) techniques. <br> MPC, the holy-grail problem of secure distributed computing, enables a set of n mutually distrusting parties to perform joint computation on their private inputs in a way that no coalition of t parties can learn more information than the output (privacy) or affect the true output of the computation (correctness). While MPC, in general, has been a subject of extensive research, the area of MPC with a small number of parties has drawn popularity of late mainly due to its application to real-time scenarios, efficiency and simplicity. This thesis focuses on designing efficient MPC frameworks for 2, 3 and 4 parties, with at most one corruption and supports ring structures. <br> Our platform aims at achieving the most substantial security notion of robustness, where the honest parties are guaranteed to obtain the output irrespective of the behaviour of the corrupt parties. A robust protocol prevents the corrupt parties from repeatedly causing the computations to rerun, thereby upholding the trust in the system. While on the roadmap to attain robustness, our frameworks also demonstrate constructions with improved performance that achieve relaxed notions of security: security with abort and fairness. A fair protocol enforces the restriction that either all parties or none of them receive the output. <br> The general structure of the computation involves the execution of the protocol steps once the participating parties have supplied their inputs. Finally, the output is distributed to all the parties. However, to enhance practical efficiency, many recent works resort to the preprocessing paradigm, which splits the computation into two phases; a preprocessing phase where input-independent (but function-dependent), computationally heavy tasks can be computed, followed by a fast online phase. Since the same functions in ML are evaluated several times, this paradigm naturally fits the case of PPML, where the ML algorithm is known beforehand. <br> At the heart of this thesis are four frameworks - ASTRA, SWIFT, Tetrad, ABY2.0 - catered to different settings. <br> - ASTRA: We begin with the setting of 3 parties (3PC), which forms the base case for honest majority. If a majority of the participating parties are honest, then the setting is deemed an honest majority setting. In the set of 3 parties, at most one party can be corrupt, and this framework tackles semi-honest corruption, where the corrupt party follows the protocol steps but tries to glean more information from the computation. ASTRA acts as a stepping stone towards achieving a stronger security guarantee against active corruption. Our protocol requires communication of 2 ring elements per multiplication gate during the online phase, attaining a per-party cost of less than one element. This is achieved for the first time in the regime of 3PC. <br> - SWIFT: Designed for 3 parties, this framework tackles one active corruption where the corrupt party can arbitrarily deviate from the computation. Building on ASTRA, SWIFT provides a multiplication that improves the communication by at least 3x over state of the art, besides improving security from abort to robustness. In the regime of malicious 3PC, SWIFT is the first robust and efficient PPML framework. It achieves a dot product protocol with communication independent of the vector size for the first time. <br> - Tetrad: Designed for 4 parties in the honest majority, the fair multiplication protocol in Tetrad requires communication of only 5 ring elements instead of 6 in the state-of-the-art. The fair framework is then extended to provide robustness without inflating the costs. A notable contribution is the design of the multiplication protocol that supports on-demand applications where the function to be computed is not known in advance. <br> - ABY2.0: Moving on to the stronger corruption model where a majority of the parties can be corrupt, we explore the base case of 2 parties (2PC). Since we aim to achieve robustness which is proven to be impossible in active corruption, we restrict ourselves to semi-honest corruption. The prime contribution of this framework is the scalar product for which the online communication is two ring elements irrespective of the vector dimension. This is a feature achieved for the first time in the 2PC literature. Along with PPML, we showcase our frameworks practicality with three relevant applications in the 2PC setting: i) AES S-box, ii) Circuit-based Private Set Intersection, iii) Biometric Matching. <br> Our frameworks provide the following contributions in addition to the ones mentioned above. First, we support multi-input multiplication for arithmetic and boolean worlds, improving the online phase in rounds and communication. Second, all our frameworks except SWIFT, incorporate truncation without incurring any overhead. Finally, we introduce efficient instantiation of garbled-world, tailor-made for the mixed-protocol framework for the first time. The mixed-protocol approach, combining arithmetic, boolean and garbled style computations, has demonstrated its potential in several practical use-cases like PPML. To facilitate the computation, we also provide the conversion mechanisms to switch between the computation styles. <br> The practicality of our framework is argued through improvements in the benchmarking of widely used ML algorithms -- Linear Regression, Logistic Regression, Neural Networks, and Support Vector Machines. One variant of our frameworks aims at minimizing the execution time, while the other focuses on the monetary cost. <br> The concrete efficiency gains of our frameworks coupled with the stronger security guarantee of robustness make our platform an ideal choice for a real-time deployment of privacy-preserving machine learning techniques. <br> <br> References: <br> [1] Harsh Chaudhari, Ashish Choudhury, Arpita Patra, Ajith Suresh. ASTRA: High Throughput 3PC over Rings with Application to Secure Prediction. ACM CCSW,19 <br> [2] Arpita Patra, Ajith Suresh. BLAZE: Blazing Fast Privacy-Preserving Machine Learning. NDSS,20 <br> [3] Harsh Chaudhari, Rahul Rachuri , Ajith Suresh. Trident: Efficient 4PC Framework for Privacy Preserving Machine Learning. NDSS20 <br> [4] Nishat Koti, Mahak Pancholi, Arpita Patra, Ajith Suresh. SWIFT: Super-fast and Robust Privacy-Preserving Machine Learning. USENIX Security,21 <br> [5] Arpita Patra, Thomas Schneider, Ajith Suresh, Hossein Yalame. ABY2.0: Improved Mixed-Protocol Secure Two-Party Computation. USENIX Security,21 <br> [6] Nishat Koti, Arpita Patra, Rahul Rachuri, Ajith Suresh. Tetrad: Fair and Robust 4PC Framework for Privacy-Preserving Machine Learning. Under Submission. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NGExMzk4NWEtMmY3My00OTViLWFlMDQtMmNlMTNhMmI2NmRj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22a787cc01-57cc-4fc1-b7e1-4e9d51923f6d%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NGExMzk4NWEtMmY3My00OTViLWFlMDQtMmNlMTNhMmI2NmRj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22a787cc01-57cc-4fc1-b7e1-4e9d51923f6d%22%7d</a> DTSTART:20210616T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210621T120000Z UID:0049e76651ac7380c9d8c9d2293dd702-159 DTSTAMP:19700101T120016Z DESCRIPTION:Specification Synthesis with Constrained Horn Clauses URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/159/specification-synthesis-with-constrained-horn-clauses/ SUMMARY:The problem of synthesizing specifications of undefined procedures has a broad range of applications, but the usefulness of the generated specifications depends on their quality. In this paper, we propose a technique for finding maximal and non-vacuous specifications. Maximality allows for more choices for implementations of undefined procedures, and non-vacuity ensures that safety assertions are reachable. To handle programs with complex control flow, our technique discovers not only specifications but also inductive invariants. Our iterative algorithm lazily generalizes non-vacuous specifications in a counterexample-guided loop. The key component of our technique is an effective non-vacuous specification synthesis algorithm. We have implemented the approach in a tool called HornSpec, taking as input systems of constrained Horn clauses. We have experimentally demonstrated the tools effectiveness, efficiency, and the quality of generated specifications, on a range of benchmarks. <br> This is joint work with Grigory Fedyukovich, Kumar Madhukar, and Deepak DSouza. The paper will be presented at the upcoming PLDI 2021 conference. <br> Google Meet joining info:<br> Video call link: <a href="https://meet.google.com/fee-usjd-oxj">https://meet.google.com/fee-usjd-oxj</a> DTSTART:20210621T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210624T120000Z UID:898bd5f98b13385a5c853f012ed3de7f-160 DTSTAMP:19700101T120015Z DESCRIPTION:Novel Reinforcement Learning Algorithms and Applications to Hybrid Control Design Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/160/novel-reinforcement-learning-algorithms-and-applications-to-hybrid-control-design-problems/ SUMMARY:The thesis is a compilation of two independent works. <br> In the first work, we develop novel weight assignment procedure, which helps us develop several schedule based algorithms.<br> Learning the value function of a given policy from the data samples is an important problem in Reinforcement Learning.<br> TD($lambda$) is a popular class of algorithms to solve this problem.<br> However, the weight assigned to different $n$-step returns decreases exponentially with increasing $n$ in TD($lambda$).<br> Here, we present a $lambda$-schedule procedure that allows flexibility in weight assignment to the different $n$-step returns.<br> Based on this procedure, we propose an on-policy algorithm, TD($lambda$)-schedule, and an off-policy algorithm, TDC($lambda$)-schedule, respectively.<br> We provide proofs of almost sure convergence for both algorithms under a general Markov noise framework as well as present the results of experiments where these algorithms are seen to show improved performance. <br> In the second work, we design hybrid control policies for hybrid systems whose mathematical models are unknown.<br> Our contributions are threefold.<br> First, we propose a framework for modelling the hybrid control design problem as a single Markov Decision Process (MDP).<br> This result facilitates the application of off-the-shelf algorithms from Reinforcement Learning (RL) literature towards designing optimal control policies.<br> Second, we model a set of benchmark examples of hybrid control design problem in the proposed MDP framework.<br> Third, we adapt the recently proposed Proximal Policy Optimisation (PPO) algorithm for the hybrid action space and apply it to the above set of problems.<br> It is observed that in each case the algorithm converges and finds the optimal policy. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NGQ2OTk0MWQtZGI5OC00OWQ1LWJmZDUtOTM4MzVhZDllNzVm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229d60e185-2600-4b28-b2d5-2d47a54928f3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NGQ2OTk0MWQtZGI5OC00OWQ1LWJmZDUtOTM4MzVhZDllNzVm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229d60e185-2600-4b28-b2d5-2d47a54928f3%22%7d</a> DTSTART:20210624T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210630T120000Z UID:a876fd0a1c47faaf476fa16d348758fa-163 DTSTAMP:19700101T120011Z DESCRIPTION:Towards Efficient Privacy-Preserving Two-Party k-Means Clustering Protocol URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/163/towards-efficient-privacy-preserving-two-party-k-means-clustering-protocol/ SUMMARY:Two-party data mining is a win-win game if played with a guarantee of data privacy from each other. This guarantee is provided by the use of cryptographic techniques in designing the two-party protocol. The need to obtain collaborative data mining results is growing and so is the need for privacy-preserving data mining protocols. Clustering is one of the data mining techniques and one of the popular clustering algorithms is k-means clustering. We studied the recent work for the secure two-party k-means clustering by Bunn and Ostrovosky and found that the protocol is inefficient for practical purposes. The protocol requires communication rounds which are linear in security parameter for the center initialization step and are quadratic in security parameter for an iterative Lloyds step of the k-means clustering algorithm. The challenge in the secure two-party k-means clustering is the exorbitant communication cost occurring due to the high number of interactions between the parties for performing computations on the data. Our work attempts to resolve this problem of inefficiency in k-means clustering protocol in a two-party setting by proposing some modifications. We have come up with two comparison protocols that are required in the k-means clustering protocol. One of the protocols is to find a minimum of two shared numbers which runs in constant communication rounds. Using this protocol as a building block, another protocol is designed to find a minimum of n shared numbers, which runs in O(n) communication rounds. We have also improved a protocol that selects a random value from a domain oblivious to both parties. Apart from this, the idea to avoid the two-party integer division altogether is incorporated in the k-means clustering protocol. With these improvements, we propose a two-party k-means clustering protocol for which the initialization step requires communication rounds linear in security parameter and Lloyds step requires communications rounds that are independent of the security parameter. The protocol provides a security guarantee in the semi-honest model except for some minor information leakage. We argue that this leakage in the protocol can be tolerated considering the substantial gain in the communication cost We have verified the gain in performance of the modified protocol by implementing both the k-means clustering protocols and running their instances in the same set-up <br> <br> Microsoft Teams link: ONLINE <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MWFjODE3MDQtNGVlYy00MmU2LTkzNzEtY2ZmNTUxYTdjMTE0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229be1cf83-a6ec-4e58-a5ca-b440a35ae199%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MWFjODE3MDQtNGVlYy00MmU2LTkzNzEtY2ZmNTUxYTdjMTE0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229be1cf83-a6ec-4e58-a5ca-b440a35ae199%22%7d</a> DTSTART:20210630T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210702T120000Z UID:5e13e47d2019c5ebc0c094eb82525efa-164 DTSTAMP:19700101T120015Z DESCRIPTION:Reinforcement Learning Algorithms for Off-Policy, Multi-Agent Learning and their Applications to Smart Grids. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/164/reinforcement-learning-algorithms-for-off-policy-multi-agent-learning-and-their-applications-to-smart-grids/ SUMMARY:Reinforcement Learning (RL) algorithms are a popular class of algorithms for training an agent to learn desired behavior through interaction with an environment whose dynamics are unknown to the agent. RL algorithms combined with neural network architectures have enjoyed much success in various disciplines like games, economics, medicine, energy management, and supply chain management. In our thesis, we study interesting extensions of single-agent RL settings, like off-policy and multi-agent settings. We discuss the motivations and importance of these settings and propose convergent algorithms to solve these problems. Finally, we consider one of the important applications of RL, namely smart grids. The goal of the smart grid is to develop an intelligent power grid model that intelligently manages its energy resources. In our thesis, we develop RL algorithms for efficient smart grid design. <br> Learning the value function of a given policy (target policy) from the data samples obtained from a different policy (behavior policy) is an important problem in Reinforcement Learning (RL). This problem is studied under the setting of off-policy prediction. Temporal Difference (TD) learning algorithms are a popular class of algorithms for solving prediction problems. TD algorithms with linear function approximation are convergent when the data samples are generated from the target policy (known as on-policy prediction) itself. However, it has been well established in the literature that off-policy TD algorithms under linear function approximation may diverge. In the first part of the thesis, we propose a convergent online off-policy TD algorithm under linear function approximation. The main idea is to penalize updates of the algorithm to ensure convergence of the iterates. We provide a convergence analysis of our algorithm. Through numerical evaluations, we further demonstrate the effectiveness of our proposed scheme. <br> Subsequently, we consider the ``off-policy`` control setup in RL, where an agents objective is to compute an optimal policy based on the data obtained from a behavior policy. As the optimal policy can be very different from the behavior policy, learning optimal behavior is very hard in the ``off-policy`` setting compared to the ``on-policy`` setting wherein the data is collected from the new policy updates. In this work, we propose the first off-policy natural actor-critic algorithm that utilizes state-action distribution correction for handling the off-policy behavior and the natural policy gradient for sample efficiency. Unlike the existing natural gradient-based actor-critic algorithms that use only fixed features for policy and value function approximation, the proposed natural actor-critic algorithm can utilize a deep neural networks power to approximate both policy and value function. We illustrate the benefit of the proposed off-policy natural gradient algorithm by comparing it with the Euclidean gradient actor-critic algorithm on benchmark RL tasks. <br> In the third part of the thesis, we consider the problem of two-player zero-sum games. In this setting, there are two agents, both of whom aim to optimize their payoffs. Both the agents observe the same state of the game, and the agents objective is to compute a strategy profile that maximizes their payoffs. However, the payoff of the second agent is the negative of the payoff obtained by the first agent. Therefore, the objective of the second agent is to minimize the total payoff obtained by the first agent.<br> This problem is formulated as a min-max Markov game in the literature. In this work, we compute the solution of the two-player zero-sum game utilizing the technique of successive relaxation. Successive relaxation has been successfully applied in the literature to compute a faster value iteration algorithm in the context of Markov Decision Processes. We extend the concept of successive relaxation to the two-player zero-sum games. We then derive a generalized minimax Q-learning algorithm that computes the optimal policy when the model information is unknown. Finally, we prove the convergence of the proposed generalized minimax Q-learning algorithm utilizing stochastic approximation techniques. Through experiments, we demonstrate the effectiveness of our proposed algorithm. <br> Next, we consider a cooperative stochastic games framework where multiple agents work towards learning optimal joint actions in an unknown environment to achieve a common goal. In many real-world applications, however, constraints are often imposed on the actions that the agents can jointly take. In such scenarios, the agents aim to learn joint actions to achieve a common goal (minimizing a specified cost function) while meeting the given constraints (specified via certain penalty functions). Our work considers the relaxation of the constrained optimization problem by constructing the Lagrangian of the cost and penalty functions. We propose a nested actor-critic solution approach to solve this relaxed problem. In this approach, an actor-critic scheme is employed to improve the policy for a given Lagrange parameter update on a faster timescale as in the classical actor-critic architecture. Using this faster timescale policy update, a meta actor-critic scheme is employed to improve the Lagrange parameters on the slower timescale. Utilizing the proposed nested actor-critic scheme, we develop three Nested Actor-Critic (N-AC) algorithms. <br> In recent times, actor-critic algorithms with attention mechanisms have been successfully applied to obtain optimal actions for RL agents in multi-agent environments. In the fifth part of our thesis, we extend this algorithm to the constrained multi-agent RL setting considered above. The idea here is that optimizing the common goal and satisfying the constraints may require different modes of attention. Thus, by incorporating different attention modes, the agents can select useful information required for optimizing the objective and satisfying the constraints separately, thereby yielding better actions. Through experiments on benchmark multi-agent environments, we show the effectiveness of our proposed algorithm. <br> In the last part of our thesis, we study the applications of RL algorithms to Smart Grids. We consider two important problems - on the supply-side and demand-side, respectively and study both in a unified framework. On the supply side, we study the problem of energy trading among microgrids to maximize profit obtained from selling power while at the same time satisfying the customer demand. On the demand side, we consider optimally scheduling the time-adjustable demand - i.e., of loads with flexible time windows in which they can be scheduled. While previous works have treated these two problems in isolation, we combine these problems and provide a unified Markov decision process (MDP) framework for these problems. <br> <br> Microsoft Teams Link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZDNjYTdhOWItYTMyMC00MzEyLWFhZWItYmQ3MzMwOTFkZmQ1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d931ce1d-4f82-48c0-8053-96272991d288%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZDNjYTdhOWItYTMyMC00MzEyLWFhZWItYmQ3MzMwOTFkZmQ1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d931ce1d-4f82-48c0-8053-96272991d288%22%7d</a> DTSTART:20210702T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210708T120000Z UID:a2aa402d88f1e708a6f23bd73526a4bf-165 DTSTAMP:19700101T120010Z DESCRIPTION:A Novel Neural Network Architecture for Sentiment-oriented Aspect-Opinion Pair Extraction URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/165/a-novel-neural-network-architecture-for-sentiment-oriented-aspect-opinion-pair-extraction/ SUMMARY:Over the years, fine-grained opinion mining in online reviews has received great attention from the NLP research community. It involves different tasks such as Aspect Term Extraction (ATE), Opinion Term Extraction (OTE), etc. Opinion Term Extraction (OTE) aims to detect different opinion expressions which convey certain attitude in the review while Aspect Term Extraction (ATE) aims to identify the entities or proposition from the review at which the attitude is directed. Recently, the NLP research community got attracted to aspect-opinion relation modeling. Such modeling would be helpful for aspect-opinion pair extraction that would be used for downstream tasks such as aspect-based sentiment analysis, opinion summarization, etc. As online reviews may contain different sentiment polarities for different aspects of the products, it would help companies find all aspects for which the customers gave positive or negative feedback. In this thesis, we propose a new opinion mining task called Sentiment-oriented Aspect-Opinion Pair Extraction (SAOPE), which aims to find all aspect-opinion pairs from customer reviews given that these pairs convey the specified sentiment polarity. We present a novel neural network architecture for the SAOPE task. In the proposed approach, aspect-opinion co-extraction is performed first and then the aspect-opinion pairs are generated through relation modeling. The aspect and the corresponding opinion words are closely related in the dependency trees. Hence, we explore graph neural networks to utilize syntactic information generated from the dependency tree of the reviews to model the relationship between the aspects and corresponding opinion words. We design a modified graph attention network (GAT) called Graph Co-attention Network (GCAT) and compare its performance with Graph Convolution Network (GCN) and Graph Attention Network (GAT) for the aspect-opinion co-extraction and the relation detection. For the SAOPE task, we evaluate our model on SemEval Challenge datasets and show that GCAT and GAT perform better than the baseline model with GCN for aspect-opinion co-extraction. We demonstrate that the proposed Graph Co-attention Network (GCAT) performs better than other graph neural networks for aspect-opinion relation detection on the publicly available benchmark datasets. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDZjMTE0ZTYtOWI5YS00MGZiLThiMjktNzM4NWZiMTNlNGEz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d DTSTART:20210708T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210706T120000Z UID:6496aa5b08c3d489f5dca89a2783caeb-166 DTSTAMP:19700101T120010Z DESCRIPTION:A Framework for Privacy Compliant Delivery Drones URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/166/a-framework-for-privacy-compliant-delivery-drones/ SUMMARY:We present Privaros, a framework to enforce privacy policies on drones. Privaros is designed for commercial delivery drones, such as the ones that will likely be used by Amazon Prime Air. Such drones visit various host airspaces, each of which may have different privacy requirements. Privaros uses mandatory access control to enforce the policies of these hosts on guest delivery drones. Privaros is tailored for ROS, a middleware popular in many drone platforms. This paper presents the design and implementation of Privaross policy-enforcement mechanisms, describes how policies are specified, and shows that policy specification can be integrated with Indias Digital Sky portal. Our evaluation shows that a drone running Privaros can robustly enforce various privacy policies specified by hosts, and that its core mechanisms only marginally increase communication latency and power consumption. <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzIwYTk3MzYtOTZjMS00Zjg4LWJjMzItN2E5MmQ4ZGQwZDJm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzIwYTk3MzYtOTZjMS00Zjg4LWJjMzItN2E5MmQ4ZGQwZDJm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d</a> DTSTART:20210706T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210712T120000Z UID:4955e85db2f38147daa98dac2da23998-167 DTSTAMP:19700101T120016Z DESCRIPTION:Automatic Code Generation for GPU Tensor Cores using MLIR URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/167/automatic-code-generation-for-gpu-tensor-cores-using-mlir/ SUMMARY:The state of the art in high-performance deep learning is primarily driven by highly tuned libraries. These libraries are often hand-optimized and tuned by expert programmers using low-level abstractions with significant effort. A lot of the effort may have to be repeated for similar hardware and future ones. Such a process is thus not modular or reusable to the same extent as compiler infrastructure like LLVM are. Manual optimization does not typically use a standard intermediate representation or transformations and passes on such intermediate representations, although the optimizations performed can be encoded as a sequence of transformation steps and customized passes. Hand tuning may also miss exploration of space or design points only reachable by automatic code generation. We believe that until the recent introduction of MLIR (Multi-level intermediate representation), intermediate representation infrastructure had not reached a stage to tackle the problem of automatic generation of libraries in a scalable and convenient manner. In particular, it was hard to represent and transform compute abstractions at high, middle and low levels using a single IR. <br> MLIR is an intermediate representation that aims to build reusable, extensible compiler infrastructure and reduce the cost of building domain-specific compilers and code generators. In this work, we tackle the problem of generating code targeting tensor cores on GPUs using the MLIR compiler infrastructure. Tensor cores are programmable matrix-multiply-and-accumulate units performing matrix-multiply accumulate operations on small matrices. First, we introduce low-level operations which are necessary to compute on tensor cores and which were absent from MLIR. Then, building on these operations, we put together a lowering pipeline that is able to fully automatically generate code for matrix-matrix multiplication (matmul) on tensor cores. Matmul is an excellent candidate to demonstrate our work as: (1) it is at the heart of many deep-learning models such as BERT, and 2) it is an excellent candidate to demonstrate various individual optimizations. We evaluate our pipeline on two different devices: 1) an NVIDIA Turing-based RTX 2080 Ti, and 2) an NVIDIA Ampere-based Geforce RTX 3090 and with two different precisions for accumulation, namely 32-bit and 16-bit wide floats. On a set of problem sizes that we evaluate, we achieve performance that is within 93% to 117% for F32 accumulate and between 79% and 158% with F16 accumulate of CuBLAS on NVIDIA Turing and Ampere respectively. We take this approach further by demonstrating the fusion of MatMul with operations that commonly follow it in deep-learning models. <br> Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZmIwODY3ODYtNmFlMC00MThmLTgwNTQtNzBjNWYyMDYwM2Jj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22171d9abc-cf43-429a-9680-c05b9523fa9a%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZmIwODY3ODYtNmFlMC00MThmLTgwNTQtNzBjNWYyMDYwM2Jj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22171d9abc-cf43-429a-9680-c05b9523fa9a%22%7d</a> DTSTART:20210712T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210713T120000Z UID:047fcc9d35d9620ce45ee46ae83c56b7-169 DTSTAMP:19700101T120016Z DESCRIPTION:Scaling Blockchains Using Coding Theory and Verifiable Computing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/169/scaling-blockchains-using-coding-theory-and-verifiable-computing/ SUMMARY:The issue of scalability has been restricting blockchains from their widespread adoption. The current transaction rate of bitcoin is around seven transactions/second while the blockchain size has crossed the 300 GB mark. Although many different approaches have been proposed to scale blockchains, e.g., sharding, lightning network, etc., we focus our analysis on methods utilizing ideas from coding theory and verifiable computing. We first consider SeF, a blockchain archiving architecture utilizing LT codes to reduce storage constraints per node up to 1000x. SeF enables full nodes to store only a small number of encoded blocks or droplets instead of an entire blockchain. Although efficient in the average case, the architecture sometimes requires large bandwidth (many droplets) to reconstruct blockchain. While other rate-less coding strategies utilizing two encoding levels are proven better than LT codes, we investigate their suitability in the proposed architecture. We propose and simulate three techniques about how to incorporate these coding strategies. The results show that precode-based rate-less coding schemes provide similar storage savings with reduced bandwidth variance for recovery. The other work we examine is PolyShard, which introduces the notion of coded-sharding. Coded sharding exports block verification of sub-ledger to the whole network instead of nodes handling that sub-ledger, making sharding resilient even to an adaptive adversary, i.e., adversary having the power to corrupt nodes after their assignment to shards. However innovative, PolyShard requires decoding of Reed-Solomon codes over large fields for block verification in real-world settings, making it computationally intensive and less practical. We propose replacing the decoding phase with verifiable computing, which reduces the bottleneck and makes the system practical for light verification functions. Microsoft teams link: https://teams.microsoft.com/l/channel/19%3axqMz7NRjdoelpofbxvfoqtfenhrRts9h0hLf8CSRIsk1%40thread.tacv2/General?groupId=68ce4fbd-dc00-4f86-98fd-48239230b6e3&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476 DTSTART:20210713T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210715T120000Z UID:c6d2910dba9d5c95612df0c07ff9da21-170 DTSTAMP:19700101T120011Z DESCRIPTION:Average Sensitivity of Graph Algorithms URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/170/average-sensitivity-of-graph-algorithms/ SUMMARY:In modern applications of graph algorithms, where the graphs of interest are large and dynamic, it is unrealistic to assume that an input representation contains the full information of a graph being studied. Hence, it is desirable to use algorithms that, even when provided with only a subgraph that misses a few edges, output solutions that are close to the solutions output when the whole graph is available. We formalize this feature by introducing the notion of average sensitivity of graph algorithms, which is the average earth movers distance between the output distributions of an algorithm on a graph and its subgraph obtained by removing an edge, where the average is over the edges removed and the distance between two outputs is the Hamming distance. <br> In this work, we initiate a systematic study of average sensitivity. After deriving basic properties of average sensitivity such as composition, we provide efficient approximation algorithms with low average sensitivities for concrete graph problems, including the minimum spanning for- est problem, the global minimum cut problem, the minimum s-t cut problem, and the maximum matching problem. One of the main ideas involved in designing our algorithms with low average sensitivity is the following fact; if the presence of a vertex or an edge in the solution output by an algorithm can be decided locally, then the algorithm has a low average sensitivity, allowing us to reuse the analyses of known sublinear-time algorithms and local computation algorithms. <br> Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20210715T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210719T120000Z UID:db9b7ae1eab6c29dd4f7abd5850df3c4-171 DTSTAMP:19700101T120012Z DESCRIPTION:Modeling and verification of database-accessing applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/171/modeling-and-verification-of-database-accessing-applications/ SUMMARY:Databases are central to the functioning of most IT-enabled processes and services. In many domains, databases are accessed and updated via applications written in general-purpose languages, as such applications need to contain the business logic and workflows that are key to the organization. Therefore, automated tools are required not only for creation and testing of database schemas and queries, etc., but also for analysis, testing, and verification of database-accessing applications. In this work we describe a novel approach for modeling, analysis and verification of database-accessing applications. We target applications that use Object Relational Mapping (ORM), which is the common database-access paradigm in most Model-View Controller (MVC) based application development frameworks. In contrast with other approaches that try to directly analyze and prove properties of complex database accessing ORM-based code, our approach infers a relational algebra specification of each controller in the application. This specification can then be fed into any off-the-shelf relational algebra solver to check properties (or assertions) given by a developer. <br> We have implemented this approach as a tool that works for Spring&quot; based MVC applications. The tool was evaluated on a set of 58 specifications. The tool found 35 of these to be satisfied; of the rest, upon manual analysis, we found that two were genuinely violated, while the remaining 21 &quot;unsatisfied&quot; warnings were actually false positives. This preliminary evaluation reveals that the approach is scalable and quite precise. <br> <br> Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NGViZTA2NDQtZjFlNi00Y2YyLTkwMGYtZmJkYTM3MjQ0N2E0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22cd42250e-1d66-4966-a431-6a8d7d5235ba%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_NGViZTA2NDQtZjFlNi00Y2YyLTkwMGYtZmJkYTM3MjQ0N2E0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22cd42250e-1d66-4966-a431-6a8d7d5235ba%22%7d</a> DTSTART:20210719T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210720T120000Z UID:2b244bd311b4374fc4fcb939686b2a3f-172 DTSTAMP:19700101T120016Z DESCRIPTION:Quantum-Safe Identity-Based Signature Scheme in Multivariate Quadratic Setting URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/172/quantum-safe-identity-based-signature-scheme-in-multivariate-quadratic-setting/ SUMMARY:Cryptographic techniques are essential for the security of communication in modern society. Today, nearly all public key cryptographic schemes used in practice are based on the two problems of factoring large integers and solving discrete logarithms. However, as the world grapples with the possibility of widespread quantum computing, these schemes are the ones most threatened. Multivariate Public Key Cryptography is one of the possible candidates for security in a post-quantum society, especially in the area of digital signature. This thesis uses the setting of multivariate cryptography to propose an identity-based signature scheme. Our proposal is based on the Rainbow signature scheme and the multivariate 3-pass identification scheme, both of which have been subjected to scrutiny by cryptographers all over the world and have emerged as strong post-quantum candidates. In our construction, we use the identity of users to generate their private key using Rainbow signature scheme. Thereafter, we use these user private keys to sign messages by applying Fiat-Shamir transform to the 3-pass identification scheme. We support the proposed scheme with suitable proof under appropriate computational assumptions, using the standard notions of security. We study the known attacks against multivariate schemes in general, and Rainbow and MQDSS in particular. We then use this analysis to propose concrete parameter sets for our construction. We implement our proposed scheme on an x86-64 PC platform and provide timing results. Our implementation shows that our construction is both practical and efficient. Thus, our proposed scheme stands as a potential post-quantum multivariate signature candidate in the identity-based setting. <br> Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZDVhMmEzZmYtY2Q2MS00ZWVhLWJiYjktZTY3ZTVhMGNiNjli%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227b997a2d-ed18-48f7-99c1-eaec40b37793%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZDVhMmEzZmYtY2Q2MS00ZWVhLWJiYjktZTY3ZTVhMGNiNjli%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227b997a2d-ed18-48f7-99c1-eaec40b37793%22%7d</a> DTSTART:20210720T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210719T120000Z UID:1ee7e9c02bc8e86583633e28025b458a-173 DTSTAMP:19700101T120016Z DESCRIPTION:Recovery Algorithms for planted structures in Semi-random models. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/173/recovery-algorithms-for-planted-structures-in-semi-random-models/ SUMMARY:For many NP-hard problems, the analysis of best-known approximation algorithms yields “poor” worst-case guarantees. However, using various heuristics, the problems can be solved (to some extent) in real-life instances. This success can be attributed to the atypicality of worst-case instances in real life, and therefore motivates studying the problem in “easier” instances. Analyzing the problem in Planted solution models and Semi-random models is one such systematic approach along these lines. <br> <br> In this thesis, we study planted solution models and semi-random models for various graph problems. Given a graph G with n vertices, we consider the task of finding the largest induced subgraph of G with a particular structure. We start by studying the problem where the particular structure is a planar graph. Next, we look at the Odd Cycle Transversal problem or equivalently the problem of finding the largest induced bipartite subgraph. Finally, we study the problem of finding the largest independent set in r-uniform hypergraphs. All these problems are NP-hard and have abysmal worst-case approximation guarantees. <br> An instance of a planted solution model is constructed by starting with a set of vertices V, and choosing a set S ⊆ V of k vertices, and adding a particular structure to it. Edges between pairs of vertices in S x (V S) and (V S) x (V S) are added independently with probability p. The algorithmic task then is to recover this planted structure. As a special case for all these problems, when the planted structure is an empty graph, the problem reduces to recovering a planted independent set and we dont expect recovery algorithms for k =o(√n). <br> For the problem of finding the largest induced bipartite subgraph, we give an exact recovery algorithm that works for k = Ωp(n log n)^1/2. For the problem of finding the maximum independent set in r-uniform hypergraphs, we give an algorithm that works for Ωp(n^{r-1/r-0.5}). Our results also hold for a natural semi-random model of instances inspired by Feige and Kilian [FK01] model. Our algorithms are based on analyzing continuous relaxations of these problems. We employ techniques from Spectral Graph Theory, Convex Optimization (Linear Programs (LPs) and Semi-Definite Programs (SDPs) relaxations), and Lasserre/Sum-of-Squares hierarchy strengthening of convex relaxations. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWJkMjBkMTEtZTQ0My00MTA3LWE1ZWYtNmM5NTM2MjY1YWI3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a2358f-870a-4a9b-aaed-f1fbcc8c2f75%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWJkMjBkMTEtZTQ0My00MTA3LWE1ZWYtNmM5NTM2MjY1YWI3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a2358f-870a-4a9b-aaed-f1fbcc8c2f75%22%7d</a> DTSTART:20210719T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210723T120000Z UID:799e0e38ff004729ad3cca329ecae140-174 DTSTAMP:19700101T120016Z DESCRIPTION:Spectral Clustering Oracles in Sublinear Time URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/174/spectral-clustering-oracles-in-sublinear-time/ SUMMARY:Given an n-vertex graph G that can be partitioned into a few clusters with good inner conductance and ϵ-sparse boundary, i.e. admits a good clustering, can we quickly tell which cluster a given vertex belongs to? A clustering oracle is a small space data structure that provides query access to an approximate clustering of the input graph in sublinear time. In this talk I will describe a clustering oracle that provides query access to an O(ϵlogk) -approximate clustering in time about n1/2+O(ϵ), where k is the number of clusters, which is essentially optimal for constant k. Our main tool is a new way of obtaining dot product access to the spectral embedding of a clusterable graph in sublinear time using the distribution of a few short random walks started at uniformly random vertices in the graph. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d DTSTART:20210723T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210723T120000Z UID:e01c6d5f09ceff7a60c3a7134b8686ad-176 DTSTAMP:19700101T120011Z DESCRIPTION:Enhancing Coverage and Robustness of Database Generators URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/176/enhancing-coverage-and-robustness-of-database-generators/ SUMMARY:Generating synthetic databases that capture essential data characteristics of client databases is a common requirement for enterprise database vendors. This need stems from a variety of use-cases, such as application testing and assessing performance impacts of planned engine upgrades. A rich body of literature exists in this area, spanning from the early techniques that simply generated data ab-initio to the contemporary ones that use a predefined client query workload to guide the data generation. In the latter category, the aim specifically is to ensure volumetric similarity -- that is, assuming a common choice of query execution plans at the client and vendor sites, the output row cardinalities of individual operators in these plans are similar in the original and synthetic databases. <br> Hydra is a recently proposed data regeneration framework that provides volumetric similarity. In addition, it also provides a mechanism to generate data dynamically during query execution, using a minuscule database summary. Notwithstanding its desirable characteristics, Hydra has the following critical limitations: (a) limited scope of SQL operators in the input query workload, (b) poor scalability with respect to the number of queries in the input workload, and (c) poor volumetric similarity on unseen queries. The data generation algorithm internally uses a linear programming (LP) solver that throttles the workload scalability. This not only puts a threshold on the training (seen) workload size but also reduces the accuracy for test (unseen) queries. Robustness towards test queries is further adversely affected by design choices such as a lack of preference among candidate synthetic databases, and artificial skew in the generated data. <br> In this work, we present an enhanced version of Hydra, called High-Fidelity Hydra (HF-Hydra), which attempts to address the above limitations. To start with, we expand the SQL operator coverage to also include the LIKE operator, and, in certain restricted settings, projection-based operators such as GROUP BY and DISTINCT. To sidestep the challenge of workload scalability, HF-Hydra outputs not one, but a suite of database summaries such that they collectively cover the entire input workload. The division of the workload into the associated sub-workloads is governed by heuristics that aim to balance robustness with LP solvability. <br> For generating richer database summaries, HF-Hydra additionally exploits metadata statistics maintained by the database engine. Further, the database query optimizer is leveraged to make the choice among the various candidate databases. The data generation is also augmented to provide greater diversity in the represented values. Finally, when a test query is fired, HF-Hydra directs it to the database summary that is expected to provide the highest volumetric similarity. <br> We have experimentally evaluated HF-Hydra on a customized set of queries based on the TPC-DS decision-support benchmark framework. We first evaluated the specialized case where each training query has its own summary, and here HF-Hydra achieves perfect volumetric similarity. Further, each summary construction took just under a second and the summary sizes were just in the order of a few tens of kilobytes. Also, our dynamic generation technique produced gigabytes of data in just a few seconds. <br> For the general setting of a limited set of summaries representing the training query workload, the data generated by HF-Hydra was compared with that from Hydra. We observed that HF-Hydra delivers more than forty percent better accuracy for outputs from filter nodes in the plans, while also achieving an improvement of about twenty percent with regard to join nodes. Further, the degradation in volumetric similarity is minor as compared to the perfectly accurate one-summary-per-query scenario, while the summary production is significantly more efficient due to reduced overheads on the LP solver. <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3azGm-zJOhD_rL-kNaZfQlJPHgDuCqDsmM8agcKyHnG2E1%40thread.tacv2/1626698346505?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a4bde2-27ad-4a57-9b14-bb2f66dfd6d0%22%7d">https://teams.microsoft.com/l/meetup-join/19%3azGm-zJOhD_rL-kNaZfQlJPHgDuCqDsmM8agcKyHnG2E1%40thread.tacv2/1626698346505?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a4bde2-27ad-4a57-9b14-bb2f66dfd6d0%22%7d</a> DTSTART:20210723T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210726T120000Z UID:99c9a43835d1bb60f14f8115432f77f3-177 DTSTAMP:19700101T120011Z DESCRIPTION:Approximation Algorithms for Geometric Packing Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/177/approximation-algorithms-for-geometric-packing-problems/ SUMMARY:We study approximation algorithms for the geometric bin packing problem and its variants. In the two-dimensional geometric bin packing problem (2D GBP), we are given n rectangular items and we have to compute an axis-parallel non-overlapping packing of the items into the minimum number of square bins of side length 1. 2D GBP is an important problem in computer science and operations research arising in logistics, resource allocation, and scheduling. We first study an extension of 2D GBP called the generalized multidimensional bin packing problem (GVBP). Here each item i additionally has d nonnegative weights v_1(i), v_2(i), …, v_d(i) associated with it. Our goal is to compute an axis-parallel non-overlapping packing of the items into bins so that for all j ∈ [d], the sum of the jth weight of items in each bin is at most 1. Despite being well studied in practice, surprisingly, approximation algorithms for this problem have rarely been explored. We first obtain two simple algorithms for GVBP having asymptotic approximation ratios (AARs) 6(d+1) and 3(1 + ln(d+1) + ε). We then extend the Round-and-Approx (R&A) framework [Bansal-Khan, SODA14] to wider classes of algorithms, and show how it can be adapted to GVBP. Using more sophisticated techniques, we obtain algorithms for GVBP having an AAR of 2(1+ln((d+4)/2))+ε, which improves to 2.919+ε for the special case of d=1. Next, we explore approximation algorithms for the d-dimensional geometric bin packing problem (dD GBP). Caprara (MOR 2008) gave a harmonic-based algorithm for dD GBP having an AAR of 1.69104^(d-1). However, their algorithm doesnt allow items to be rotated. This is in contrast to some common applications of dD GBP, like packing boxes into shipping containers. We give approximation algorithms for dD GBP when items can be orthogonally rotated about all or a subset of axes. We first give a fast and simple harmonic-based algorithm, called fullh_k, having an AAR of 1.69104^d. We next give a more sophisticated harmonic-based algorithm, which we call hgap_k, having an AAR of (1+ε)1.69104^(d-1). This gives an AAR of roughly 2.860 + ε for 3D GBP with rotations, which improves upon the best-known AAR of 4.5. In addition, we study the multiple-choice bin packing problem that generalizes the rotational case. Here we are given n sets of d-dimensional cuboidal items and we have to choose exactly one item from each set and then pack the chosen items. Our algorithms fullh_k and hgap_k also work for the multiple-choice bin packing problem. We also give fast and simple approximation algorithms for the multiple-choice versions of dD strip packing and dD geometric knapsack. These algorithms have AARs 1.69104^(d-1) and (1-ε)3^(-d), respectively. A rectangle is said to be δ-skewed if it has width at most δ or height at most δ. We give an approximation algorithm for bin packing δ-skewed rectangles whose asymptotic approximation ratio approaches 1 as δ approaches 0. Our result indicates that hard instances in geometric bin packing arise due to items that are large in both dimensions. A packing of rectangles into a bin is said to be guillotine-separable iff we can use a sequence of end-to-end cuts to separate the items from each other. The asymptotic price of guillotinability (APoG) is the maximum value of opt_G(I)/opt(I) for large opt(I), where opt(I) and opt_G(I) are the minimum number of bins and the minimum number of guillotine-separable bins, respectively, needed to pack I. Computing lower and upper bounds on APoG is an important problem, since proving an upper bound smaller than 1.5 would beat the state-of-the-art algorithm for 2D GBP. The best-known lower and upper bounds are 4/3 and 1.69104, respectively. We analyze this problem for the special case of δ-skewed rectangles, where δ is a small constant (i.e., close to 0). We give a roughly 4/3-asymptotic-approximate algorithm for 2D GBP for this case, and our algorithms output is guillotine-separable. This proves an upper-bound of roughly 4/3 on APoG for δ-skewed rectangles. We also prove a matching lower-bound of 4/3. This shows that hard examples for upper-bounding APoG include items that are large in both dimensions. Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_NWU4MzkzMWEtZWY0ZS00MjU3LTk1YzMtMTMxZDVmZjM5Nzhl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22cd1ddf68-b75e-4337-87f3-ded65154fa20%22%7d DTSTART:20210726T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210728T120000Z UID:478474418b64ee5ff6818d9b8e0bf9f7-178 DTSTAMP:19700101T120016Z DESCRIPTION:Hadwigers conjecture and total coloring URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/178/hadwigers-conjecture-and-total-coloring/ SUMMARY:Hadwigers conjecture is one of the most important and long-standing conjectures in graph theory. It was established that proving Hadwigers conjecture is equivalent to proving the conjecture for the special class of graphs that can be expressed as the square of some other graph. However, it is difficult to prove Hadwigers conjecture even for the squares of highly specialised graph classes. We decided to investigate the squares of subdivided graphs (a subdivided graph is a graph that can be obtained from another graph $H$ by replacing each edge $uv$ of $H$ by a path $uwv$, where $w$ is a new vertex). <br> It turns out that squares of subdivided graphs are exactly the class of total graphs, well-studied in the context of the total coloring conjecture, another well-known and long-standing conjecture in graph theory. The total graph of $G$, denoted by $T(G)$, is defined on the vertex set $V(G)sqcup E(G)$ with $c_{1},c_{2}in V(G)sqcup E(G)$ adjacent whenever $c_{1}$ and $c_{2}$ are adjacent to or incident on each other in $G$. The total-chromatic number $chi(G)$ of a graph $G$ is defined to be equal to the chromatic number of its total graph. That is, $chi(G)=chi(T(G))$. <br> The total coloring conjecture or textit{TCC} states that for every graph $G$, $chi(G)leqDelta(G)+2$. Though Hadwigers conjecture was proved for the closely related class of line graphs 20 years ago itself, Hadwigers conjecture is not yet studied in the context of total graphs. In this thesis, the following results are proved: (i) There exists a constant $C$ such that, if the connectivity of $Ggeq C$, then Hadwigers conjecture is true for $T(G)$. <br> (ii) Let $mathcal{F}$ be a class of graphs that is closed under the operation of taking subgraphs. If a weaker version of the total coloring conjecture (weak textit{TCC}) namely, $chi(G)leqDelta(G)+3$, is true for the class $mathcal{F}$, then Hadwigers conjecture is true for the class ${T(G): Gin mathcal{F} }$. The second statement motivates one to look for classes of graphs that satisfy weak textit{TCC}. It may be noted that a complete proof of textit{TCC} for even $4$-colorable graphs (in fact even for planar graphs) has remained elusive even after decades of effort; but weak textit{TCC} can be proved easily for $4$-colorable graphs. <br> It was noticed that in spite of interest in studying $chi(G)$ in terms of $chi(G)$ right from the initial days, weak textit{TCC} is not proven to be true for $k$-colorable graphs even for $k=5$.<br> In the latter half of the thesis, an important contribution to the total coloring literature is made by proving that $chi(G)leq Delta(G)+3$, for every $5$-colorable graph $G$. Being close to some of the well studied topics in total coloring, it seems that this is a long-pending addition to the literature. <br> Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGMzMDhiNTItM2IwNS00MDZhLWI4Y2ItMThhMjkwN2M4YjNi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22ed2e5ccd-b870-455c-862e-26e9ab1908be%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGMzMDhiNTItM2IwNS00MDZhLWI4Y2ItMThhMjkwN2M4YjNi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22ed2e5ccd-b870-455c-862e-26e9ab1908be%22%7d</a> DTSTART:20210728T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210727T120000Z UID:679f0763e42ce79273be2c4118376dea-180 DTSTAMP:19700101T120009Z DESCRIPTION:Locally Reconstructable Non-Malleable Secret Sharing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/180/locally-reconstructable-non-malleable-secret-sharing/ SUMMARY:Non-malleable secret sharing (NMSS) schemes, introduced by Goyal and Kumar (STOC 2018), ensure that a secret m can be distributed into shares m1,...,mn (for some n), such that any t (a parameter <= n) shares can be reconstructed to recover the secret m, any t-1 shares doesnt leak information about m and even if the shares that are used for reconstruction are tampered, it is guaranteed that the reconstruction of these tampered shares will either result in the original m or something independent of m. Since their introduction, non-malleable secret sharing schemes sparked a very impressive line of research. <br> In this talk, we present a new feature of local reconstructablility in NMSS, which allows reconstruction of any portion of a secret by reading just a few locations of the shares. This is a useful feature, especially when the secret is long or when the shares are stored in a distributed manner on a communication network. In this talk, we give a compiler that takes in any non-malleable secret sharing scheme and compiles it into a locally reconstructable non-malleable secret sharing scheme. To secret share a message consisting of k blocks of length r each, our scheme would only require reading r + log k bits (in addition to a few more bits, whose quantity is independent of r and k) from each partys share (of a reconstruction set) to locally reconstruct a single block of the message. <br> Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzNjOGVkNzgtNTM5Zi00OWRlLWJkYWItYWNjMjI4ZGI1NzJi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%225f3273b8-8838-46b7-b675-b1e9eab4d8ef%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzNjOGVkNzgtNTM5Zi00OWRlLWJkYWItYWNjMjI4ZGI1NzJi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%225f3273b8-8838-46b7-b675-b1e9eab4d8ef%22%7d</a> DTSTART:20210727T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210730T120000Z UID:1926a4477be6b260f79220ed136b854b-181 DTSTAMP:19700101T120016Z DESCRIPTION:Superpolynomial lower bounds against low-depth algebraic circuits URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/181/superpolynomial-lower-bounds-against-low-depth-algebraic-circuits/ SUMMARY:An algebraic circuit computes a polynomial using addition and multiplication operators. Understanding the power of algebraic circuits has close connections to understanding general computation. It is known that proving lower bounds for algebraic circuits can serve as a stepping stone towards proving general Boolean circuit lower bounds. <br> Despite this, not many lower bounds are known for even simple Sigma Pi Sigma (product-depth 1) circuits. Before our work, the best known lower bound for product-depth 1 circuit was (slightly less than) cubic. No lower bounds were known for general product-depth 2 circuits. <br> In this work, we show the first superpolynomial lower bound for low-product-depth algebraic circuits. <br> In the talk, we discuss the main results and present the proof ideas used in the proof of the superpolynomial lower bound for product-depth 1 circuits. <br> This talk is based on joint work with Srikanth Srinivasan and Sébastien Tavenas. <br> Microsoft Teams Link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20210730T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210728T120000Z UID:b8e3b0c3467fd915e31999504dae31d3-182 DTSTAMP:19700101T120015Z DESCRIPTION:Algorithms for Fair Clustering URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/182/algorithms-for-fair-clustering/ SUMMARY:Many decisions today are taken by various machine learning algorithms, hence it is crucial to accommodate fairness in such algorithms to remove/reduce any kind of bias in the decision. We incorporate fairness in the problem of clustering. Clustering is a classical machine learning problem in which the task is to partition the data points into various groups such that the data points belonging to one group are more similar to each other than the data points belonging to some other group in the partition.<br> In our model, each data point belongs to one or more number of categories. We define fairness in terms of two constraints, restricted dominance and minority protection. While ensuring fairness in the clustering, we consider each data point in only of the categories from the set of categories it belongs to. Our model ensures that no category is either in minority or in dominance in any of the clusters. Representation of a category in a cluster is considered not in absolute terms but in proportion to its presence in the whole dataset. We give bi-criteria approximation for fair clustering whose objective is to minimise $L_p$-norm where $p$ can take any integral value. We implement this algorithm and do experiments to compare it with the state-of-the-art. For any $epsilon >0$, we give a randomized algorithm with approximation ratio of $(1 + epsilon)$ for fair clustering for points lying in Euclidean space whose objective is to minimise $L_1$-norm (or $L_2$-norm). For points lying in $mathbb{R}^d$, the run time of this algorithm for $L_2$-norm is $Oleft(nd cdot 2^{tilde{O}(k/epsilon)}right) + poly(n) cdot 2^{tilde{O}(k/epsilon)}$, where $n$ represents the size of the dataset. For $L_1$-norm, the run time of this algorithm is $Oleft(nd cdot 2^{tilde{O}(k/epsilon^{O(1)})}right) + poly(n) cdot 2^{tilde{O}(k/epsilon^{O(1)})} $. Given a $gamma$-perturbation resilient instance of clustering in the metric space $(V,d)$, we also give a bi-criteria approximation for the fair clustering of the same instance while changing its metric to $d $. Here, $d $ is any metric which is a $gamma$-perturbation of $(V,d)$. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjZkYWU2MjEtMGYwZS00YmZlLWI1MGMtMzA3YjhiMGUyZjgy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220c3c2a63-37e3-4ad6-b0bd-ddfa9589e2d5%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjZkYWU2MjEtMGYwZS00YmZlLWI1MGMtMzA3YjhiMGUyZjgy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220c3c2a63-37e3-4ad6-b0bd-ddfa9589e2d5%22%7d</a> DTSTART:20210728T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210730T120000Z UID:b6c69b0de1e8a14e1a1cc7d28435df5f-183 DTSTAMP:19700101T120014Z DESCRIPTION:Network Anonymity, Privacy, (Anti-) Censorship and the Whole Nine Yards. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/183/network-anonymity-privacy-anti-censorship-and-the-whole-nine-yards/ SUMMARY:In the second decade of the century (circa the Arab Springs of 2011), the Internet is the new battlefield where wars between politicians, media, (h)activists, lawyers and the military, shape the destiny of millions of people. Historically incepted as the ARPANET, it was engineered to serve as means of communication, even in the face of calamities and wars. Political will often plays antithetical to this very attribute. For instance, countries like China, Iran and UAE use (homebrewed) firewalling infrastructure to censor web traffic -- sometimes with the pretext of preserving cultural and religious values, at other times to prevent political dissent. No wonder a large body of network censorship measurements focuses on these two countries. While such countries are inherently (constitutionally) undemocratic, free speech over the Internet is, in recent years, being regularly suppressed even in democracies like India. Such evolutions are positioned on concerns otherwise paramount to the preservation of human rights -- e.g., policing child pornography. But state control of communication channels has been abuse to silence dissent, even in India where the supreme court deems freedom of speech on the Internet a fundamental right. <br> In this context, it is natural to ask how free and open is the Internet and how robust it is to censorship by countries like India, that in the recent years has evolved a sophisticated censorship infrastructure. <br> In this talk I present an overview our work over the years that has focussed on evolution of Indians Internet censorship infrastructure, how it censors traffic (and now apps.), how various ISPs implement it. Further, I also present some research efforts to evade censorship (and also Internet shutdowns/blackouts). <br> To begin with we consider the question of whether India might potentially follow the Chinese model and institute a single, government-controlled filter. Our research shows that would not be difficult, as the Indian Internet is quite centralized already. A few key ASes (~ 1% of Indian ASes, i.e. less than 4) and routers (<5000) collectively intercept approximately 95% of paths to the censored sites and to all publicly-visible DNS resolvers. Thereafter we conducted an extensive study (first of its kind) involving nine major ISPs of the country in terms of what kind of censorship techniques they use, what triggers them, their consistency and coverage, and how to evade them. Our results indicate a clear disparity among the ISPs, on how widely they install censorship infrastructure. As of 2021, we have extensively explored the evolution of web censorship (HTTPS) along with exactly how Chinese apps are being filtered in the country. While existing solutions to evade censorship include proxies, VPNs, Tor have been designed primarily for web, while other applications like VoIP (real-time voice) are mostly ignored. As a part of our research we have extensively explored the feasibility of transporting real-time voice (mostly UDP) over Tor (that primarily supports TCP). Prior research deemed Tor to be unsuitable for such purposes. In our research we tried to identify how the interplay of network attributes (delay, jitter, bandwidth etc.) impact performance of VoIP. To our surprise the belief established from prior research seems unfounded. <br> However, all such solutions that rely on proxies are prone to being filtered by the ISPs, as these end-points are easily discoverable. Futuristic solutions like Decoy Routing, that rely on routers that could double as “smart proxies”, are resilient to such filtering. They have hitherto relied mostly on commodity servers, and involve wide scale traffic observation, inadvertently posing a threat to the privacy of users who do not require such services. To that end, we devised a SDN based DR solution, SiegeBreaker, that not only performs at line rates (comparable to native TCP) but also does not require inspection of all network flows, thus preserving the privacy of oblivious users. However, the deployability of such solutions remains a challenge, as it requires support from major top-tier ISPs. <br> A third alternative, combining the best of both the above solutions, involves tunnelling Internet traffic over that of various (semi-)real time applications, e.g. Instant Messaging (IM). To that end, we designed and tested a scheme, Camoufler, that utilizes IM channels as-is for transporting traffic. The scheme provides unobservability and good QoS, due to its inherent properties, such as low-latency message transports. Moreover, unlike Decoy Routing, it does not pose new deployment challenges. Performance evaluation of Camoufler, implemented on five popular IM apps indicate that it provides sufficient QoS for web browsing. E.g., the median time to render the homepages of Alexa top-1k sites was recorded to be about 3.6s, when using Camoufler implemented over Signal. <br> Finally, I would like to conclude the talk with our new system Dolphin, that emulates old school dial-up modems, sans the ISP support, to relay Internet traffic especially in the face of Internet shutdowns. Dolphins protocol recovers from the losses and errors introduced by the cellular voice medium, while also assuring end-to-end confidentiality. At low data rates (<=64bps), the errors are under 5% and suitable for supporting delay-tolerant applications with acceptable latencies. E.g. a 280 character tweet can be posted in about a minute. DTSTART:20210730T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210817T120000Z UID:af47918db357dd6b4a86d51ce183efa2-184 DTSTAMP:19700101T120014Z DESCRIPTION:nuKSM: NUMA-aware Memory De-duplication for Multi-socket Servers URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/184/nuksm-numa-aware-memory-de-duplication-for-multi-socket-servers/ SUMMARY:Teams meeting link: <a href="shorturl.at/jsAR9">shorturl.at/jsAR9 </a> <br> <br> Memory management is one of the most critical pieces in an operating system's design. It has several responsibilities ranging from ensuring quick access to data by applications to enabling memory consolidation. For example, judicious placement of pages in multi-socket NUMA (non-uniform memory access) servers could determine the access latencies experienced by an application. Similarly, memory de-duplication can play a pivotal role in memory consolidation and over-commitment. <br> Different responsibilities of memory management can conflict with each other. This often happens when different subsystems of an OS are responsible for different memory management goals, and each works in its silo. In this work, we report one such conflict that appears between memory de-duplication and NUMA management. Linux's memory de-duplication subsystem, namely KSM, is NUMA unaware. We demonstrate that memory de-duplication can have unintended consequences to NUMA overheads experienced by applications running on multi-socket servers. Linux's memory de-duplication subsystem, namely KSM, is NUMA unaware. Consequently, while de-duplicating pages across NUMA nodes, it can place de-duplicated pages in a manner that can lead to significant performance variations, unfairness, and subvert process priority. <br> We introduce NUMA-aware KSM, a.k.a., nuKSM, that makes judicious decisions about the placement of de-duplicated pages to reduce the impact of NUMA and unfairness in execution. nuKSM also enables users to avoid priority subversion. Finally, independent of the NUMA effect, we observed that KSM fails to scale well to large memory systems due to its centralized design. We thus extended nuKSM to adopt a decentralized design to scale to larger memory. DTSTART:20210817T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210806T120000Z UID:093a8640324444248abc8d401f1948fb-185 DTSTAMP:19700101T120011Z DESCRIPTION:Recent progress in online matching - General arrival models URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/185/recent-progress-in-online-matching-general-arrival-models/ SUMMARY:Originated from the seminal work by Karp, Vazirani, and Vazirani (1990), online matching has been established as one of the most fundamental topics in the literature of online algorithms. <br> This is the first of two talks that presents the basics of online matching, and surveys the recent progress in two directions. Todays talk will be focussing on General arrival models and for the next talk, we will look at some Open problems in online advertising. <br> General arrival models <br> Traditional online matching models consider bipartite graphs and assume knowing one side of the bipartite graph upfront. The matching problems in many modern scenarios, however, do not fit into the traditional models. In the problem of matching ride-sharing requests, for instance, the graph is not bipartite in general, and all vertices arrive online. There has been much progress in the past three years on online matching models beyond the traditional ones, including the fully online model, the general vertex arrival model, and the edge arrival model. <br> For more details about the seminar please visit the website at <a href="https://www.csa.iisc.ac.in/iisc-msr-seminar/">https://www.csa.iisc.ac.in/iisc-msr-seminar/</a> <br> Microsoft Teams Link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20210806T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210813T120000Z UID:25f97ca767b4fe8dac5f97b021b7a969-186 DTSTAMP:19700101T120011Z DESCRIPTION:Recent progress in online matching - Some open problems in online advertising URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/186/recent-progress-in-online-matching-some-open-problems-in-online-advertising/ SUMMARY:Originated from the seminal work by Karp, Vazirani, and Vazirani (1990), online matching has been established as one of the most fundamental topics in the literature of online algorithms. <br> This is the second of two talks which presents the basics of online matching. The first talk focussed on General arrival models and todays talk will be looking at some Open problems in online advertising. <br> Open problems in online advertising: <br> AdWords and Display Ads are generalizations of the online bipartite matching problem by Karp et al. These problems capture online advertising which generates tens of billions of US dollars annually. This year, we introduce a new technique called online correlated selection, and design the first online algorithms for the general cases of AdWords and Display Ads outperforming greedy, which has remained the state of the art for more than 10 years, despite many attempts to find better alternatives. <br> For more details about the seminar please visit the website at <a href="https://www.csa.iisc.ac.in/iisc-msr-seminar/">https://www.csa.iisc.ac.in/iisc-msr-seminar/</a> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20210813T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210809T120000Z UID:4421b7d6460c3b852988344a4a22cbc8-187 DTSTAMP:19700101T120014Z DESCRIPTION:Near-Optimal Non-malleable Codes and Leakage Resilient Secret Sharing Schemes URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/187/near-optimal-non-malleable-codes-and-leakage-resilient-secret-sharing-schemes/ SUMMARY:Non-malleable codes (NMCs) are coding schemes that help in protecting crypto-systems under tampering attacks, where the adversary tampers the device storing the secret and observes additional input-output behavior on the crypto-system. NMCs give a guarantee that such adversarial tampering of the encoding of the secret will lead to a tampered secret, which is either same as the original or completely independent of it, thus giving no additional information to the adversary. Leakage resilient secret sharing schemes help a party, called a dealer, to share his secret message amongst n parties in such a way that any t of these parties can combine their shares to recover the secret, but the secret remains hidden from an adversary corrupting < t parties to get their complete shares and additionally getting some bounded bits of leakage from the shares of the remaining parties. <br> For both these primitives, whether you store the non-malleable encoding of a message on some tamper-prone system or the parties store shares of the secret on a leakage-prone system, it is important to build schemes that output codewords/shares that are of optimal length and do not introduce too much redundancy into the codewords/shares. This is, in particular, captured by the rate of the schemes, which is the ratio of the message length to the codeword length/largest share length. The research goal of the thesis is to improve the state of art on rates of these schemes and get near-optimal/optimal rates. <br> In this talk, I will specifically focus on leakage resilient secret sharing schemes, describe the leakage model, and take you through the state of the art on their rates. Finally, I will present a recent construction of an optimal (constant) rate, leakage resilient secret sharing scheme in the so-called "joint and adaptive leakage model" where leakage queries can be made adaptively and jointly on multiple shares. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/dl/launcher/launcher.html?url=%2F_%23%2Fl%2Fmeetup-join%2F19%3Ameeting_ZjBiZDE0ZjgtYzM3Mi00YmQ4LTllZjMtZGI4YjBlNDdkMmI2%40thread.v2%2F0%3Fcontext%3D%257b%2522Tid%2522%253a%25226f15cd97-f6a7-41e3-b2c5-ad4193976476%2522%252c%2522Oid%2522%253a%25220144d79c-31b7-4e38-a5c7-4aacd1276766%2522%257d%26anon%3Dtrue&type=meetup-join&deeplinkId=0be388dc-57e8-4798-bb75-8e604af5b680&directDl=true&msLaunch=true&enableMobilePage=true&suppressPrompt=true">https://teams.microsoft.com/dl/launcher/launcher.html?url=%2F_%23%2Fl%2Fmeetup-join%2F19%3Ameeting_ZjBiZDE0ZjgtYzM3Mi00YmQ4LTllZjMtZGI4YjBlNDdkMmI2%40thread.v2%2F0%3Fcontext%3D%257b%2522Tid%2522%253a%25226f15cd97-f6a7-41e3-b2c5-ad4193976476%2522%252c%2522Oid%2522%253a%25220144d79c-31b7-4e38-a5c7-4aacd1276766%2522%257d%26anon%3Dtrue&type=meetup-join&deeplinkId=0be388dc-57e8-4798-bb75-8e604af5b680&directDl=true&msLaunch=true&enableMobilePage=true&suppressPrompt=true</a> DTSTART:20210809T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210811T120000Z UID:f7a904e7cd1223e9480f467f7f837749-188 DTSTAMP:19700101T120017Z DESCRIPTION:What makes a good applied Data scientist? URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/188/what-makes-a-good-applied-data-scientist/ SUMMARY:In this talk, I will briefly cover the rise of Data Science as an applied field. I will borrow examples from e-commerce to demonstrate how Data Science gets used in real life. I will discuss a few recent trends in data science and how development life cycles are changing. Finally, I will share a few tips on what makes a good applied data scientist. I will make a case for a well rounded computer science education (sound fundamentals of computer systems, algorithm analysis and design) along with some cross disciplinary exposure as essential to succeed in this rapidly evolving area. There will be plenty of room for questions and looking forward to an interactive discussion. DTSTART:20210811T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210809T120000Z UID:98265327a962cc484ad6f238a98096e8-189 DTSTAMP:19700101T120014Z DESCRIPTION:Near-Optimal Non-malleable Codes and Leakage Resilient Secret Sharing Schemes URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/189/near-optimal-non-malleable-codes-and-leakage-resilient-secret-sharing-schemes/ SUMMARY:Non-malleable codes (NMCs) are coding schemes that help in protecting crypto-systems under tampering attacks, where the adversary tampers the device storing the secret and observes additional input-output behavior on the crypto-system. NMCs give a guarantee that such adversarial tampering of the encoding of the secret will lead to a tampered secret, which is either same as the original or completely independent of it, thus giving no additional information to the adversary. Leakage resilient secret sharing schemes help a party, called a dealer, to share his secret message amongst n parties in such a way that any t of these parties can combine their shares to recover the secret, but the secret remains hidden from an adversary corrupting < t parties to get their complete shares and additionally getting some bounded bits of leakage from the shares of the remaining parties. <br> For both these primitives, whether you store the non-malleable encoding of a message on some tamper-prone system or the parties store shares of the secret on a leakage-prone system, it is important to build schemes that output codewords/shares that are of optimal length and do not introduce too much redundancy into the codewords/shares. This is, in particular, captured by the rate of the schemes, which is the ratio of the message length to the codeword length/largest share length. The research goal of the thesis is to improve the state of art on rates of these schemes and get near-optimal/optimal rates. <br> In this talk, I will specifically focus on leakage resilient secret sharing schemes, describe the leakage model, and take you through the state of the art on their rates. Finally, I will present a recent construction of an optimal (constant) rate, leakage resilient secret sharing scheme in the so-called DTSTART:20210809T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210821T120000Z UID:5c71185a740820e753eb4f4cd0d74218-190 DTSTAMP:19700101T120017Z DESCRIPTION:The Story of Arjun Guha, or: the Arc of a Research Project URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/190/the-story-of-arjun-guha-or-the-arc-of-a-research-project/ SUMMARY:There are many ways to do research successfully, so you should take all advice with a pinch of salt. That said, I will try to describe some of the attributes that have worked well for my students and me. To make it concrete, I will use my former student, Arjun Guha, as a case study. I will strip away most of the technical detail while trying to retain the essential ideas, so people from any area of computer science will be able to follow almost all of it. Questions are welcome at any time. The talk will be followed by plenty of time for Q&A. DTSTART:20210821T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210820T120000Z UID:eb1e849ed849bdf0c7413c0604cf4b1e-191 DTSTAMP:19700101T120016Z DESCRIPTION:A 3-Approximation Algorithm for Maximum Independent Set of Rectangles URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/191/a-3-approximation-algorithm-for-maximum-independent-set-of-rectangles/ SUMMARY:We study the Maximum Independent Set of Rectangles (MISR) problem, where we are given a set of axis-parallel rectangles in the plane and the goal is to select a subset of non-overlapping rectangles of maximum cardinality. In a recent breakthrough, Mitchell [2021] obtained the first constant-factor approximation algorithm for MISR. His algorithm achieves an approximation ratio of 10 and it is based on a dynamic program that intuitively recursively partitions the input plane into special polygons called corner-clipped rectangles, without intersecting certain special horizontal line segments called fences. <br> <br> In this work, we present a 3-approximation algorithm for MISR which is based on a similar recursive partitioning scheme. First, we use a partition into a more general class of axis-parallel polygons with constant complexity each, which allows us to provide an arguably simpler analysis and at the same time already improves the approximation ratio to 6. Then, using a more elaborate charging scheme and a recursive partitioning into general axis-parallel polygons with constant complexity, we improve our approximation ratio to 3. In particular, our partitioning uses more general fences that can be sequences of up to O(1) line segments each. This and our other new ideas may be useful for future work towards a PTAS for MISR. <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20210820T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210826T120000Z UID:bc16ce5a4d67359b7902b10879ebbb02-192 DTSTAMP:19700101T120014Z DESCRIPTION:Neural Models for Personalized Recommendation Systems with External Information URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/192/neural-models-for-personalized-recommendation-systems-with-external-information/ SUMMARY:Personalized recommendation systems use the data generated by user-item interactions (for example, in the form of ratings) to predict different users interests in available items and recommend a set of items or products to the users. The sparsity of data, cold start, and scalability are some of the important challenges faced by the developers of recommendation systems. These problems are alleviated by using external information, which can be in the form of a social network or a heterogeneous information network, or cross-domain knowledge. This thesis develops novel neural network models for designing personalized recommendation systems using the available external information. <br> The first part of the thesis studies the top-N item recommendation setting where the external information is available in the form of a social network or heterogeneous information network. Unlike a simple recommendation setting, capturing complex relationships amongst entities (users, items, and connected objects) becomes essential when a social and heterogeneous information network is available. In a social network, all socially connected users do not have equal influence on each other. Further, estimating the quantum of influence among entities in a user-item interaction network is important when only implicit ratings are available. We address these challenges by proposing a novel neural network model, SoRecGAT, which employs a multi-head and multi-layer graph attention mechanism. The attention mechanism helps the model learn the influence of entities on each other more accurately. Further, we exploit heterogeneous information networks (HIN) to gather multiple views for the items. A novel neural network model -- GAMMA (Graph and Multi-view Memory Attention mechanism) is proposed to extract relevant information from HINs. The proposed model is an end-to-end model which eliminates the need for learning a similarity matrix offline using some manually selected meta-paths before optimizing the desired objective function. <br> In the second part of the thesis, we focus on top-N bundle recommendation and list continuation setting. Bundle recommendation is the task of recommending a group of products instead of individual products to users. We study two interesting challenges -- (1) how to personalize and recommend existing bundles to users and (2) how to generate personalized novel bundles targeting specific users. We propose GRAM-SMOT -- a graph attention-based framework that considers higher-order relationships among the users, items, and bundles and the relative influence of items present in the bundles. For efficiently learning the embeddings of the entities, we define a loss function based on the metric-learning approach. A strategy that leverages submodular optimization ideas is used to generate novel bundles. <br> We also study the problem of top-N personalized list continuation where the task is to curate the next items to user-generated lists (ordered sequence of items) in a personalized way by using the sequential information of the items in the list. The main challenge in this task is understanding the ternary relationships among the users, items, and lists. We propose HyperTeNet -- a self-attention hypergraph and Transformer-based neural network architecture for the personalized list continuation task. Here, graph convolutions are used to learn the multi-hop relationship among entities of the same type. A self-attention-based hypergraph neural network is proposed to learn the ternary relationships among the interacting entities via hyperlink prediction in a 3-uniform hypergraph. Further, the entity embeddings are shared with a Transformer-based architecture and are learned through an alternating optimization procedure. <br> The final part of the thesis focuses on the personalized rating prediction setting where external information is available in the form of cross-domain knowledge. We propose an end-to-end neural network model, NeuCDCF, that provides a way to alleviate data sparsity problems by exploiting the information from related domains. NeuCDCF is based on a wide and deep framework and learns the representations jointly using matrix factorization and deep neural networks. We study the challenges involved in handling diversity between domains and learning complex non-linear relationships among entities within and across domains. <br> We conduct extensive experiments in each of these settings using several real-world datasets and demonstrate the efficacy of the proposed models. <br> <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MWVhYWRkODUtZGI4OC00ZDZmLTk2NGMtZmViNTNiZmFmM2E3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MWVhYWRkODUtZGI4OC00ZDZmLTk2NGMtZmViNTNiZmFmM2E3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20210826T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210903T120000Z UID:408b362279c30f9cecc0f7517b5aeed8-193 DTSTAMP:19700101T120016Z DESCRIPTION:A (2 + ϵ)-approximation algorithm for preemptive weighted flow time on a single machine URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/193/a-2-i%c2%b5-approximation-algorithm-for-preemptive-weighted-flow-time-on-a-single-machine/ SUMMARY:In a recent breakthrough in scheduling, Batra, Garg, and Kumar gave the first constant approximation algorithm for minimizing the sum of weighted flow times. Wiese and I (STOC 21) managed to improve this large unspecified constant to 2+ϵ. I will give a very graphic presentation of the algorithmic techniques behind this. <br> <br> Microsoft Teams Link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> <br> For more details about the seminar please visit the website at <a href="https://www.csa.iisc.ac.in/iisc-msr-seminar/">https://www.csa.iisc.ac.in/iisc-msr-seminar/</a> DTSTART:20210903T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210903T120000Z UID:03e906d83e472a857fc2281cf32d0733-194 DTSTAMP:19700101T120014Z DESCRIPTION:A Syntactic Neural Model for Question Decomposition URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/194/a-syntactic-neural-model-for-question-decomposition/ SUMMARY:Question decomposition along with single-hop Question Answering (QA) system serve as useful modules in developing multi-hop Question Answering systems, mainly because the resulting QA system is interpretable and has been demonstrated to exhibit better performance. The problem of Question Decomposition can be posed as a machine translation problem and it can be solved using any sequence-to-sequence neural architecture. Using this approach, it is difficult to capture the innate hierarchical structure of the decomposition. Inspired by database query languages a pseudo-formalism for capturing the meaning of questions, called Question Decomposition Meaning Representation (QDMR) was recently introduced. In this approach, a complex question is decomposed into simple queries which are mapped into a small set of formal operations. This method does not utilize the underlying syntax information of QDMR to generate the decomposition. <br> In the area of programming language code generation, methods that use syntax information as a prior knowledge have been demonstrated to perform better. Moreover, the syntax-aware models are usually interpretable. Motivated by the success of syntax-aware models, we propose a new syntactic neural model for question decomposition in this thesis. In particular, we encode the underlying syntax of the QDMR structures into a grammar model as a sequence of actions. <br> This is done using a deterministic framework which uses Abstract Syntax Trees (AST) and Parse Trees. The proposed approach can be thought of as an encoder-decoder method for QDMR structures where a sequence of possible actions is a latent representation of the QDMR structure. The advantage of using this latent representation is that it is interpretable. Experimental results on a real-world dataset demonstrate that the proposed approach outperforms the state-of-the-art approach especially in scenarios where training data is limited. Some heuristics to further improve the performance of the proposed approach are also suggested in this work. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YTQ3YzdhOWMtMDBiYy00ODQxLWJmMDItMzlmY2MwNjFjYjY5%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YTQ3YzdhOWMtMDBiYy00ODQxLWJmMDItMzlmY2MwNjFjYjY5%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20210903T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210910T120000Z UID:c18547618ffed2f9ab0d4291b0044971-195 DTSTAMP:19700101T120016Z DESCRIPTION:Better-Than-2 Approximations for Weighted Tree Augmentation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/195/better-than-2-approximations-for-weighted-tree-augmentation/ SUMMARY:The Weighted Tree Augmentation Problem (WTAP) is one of the most basic connectivity augmentation problems. It asks how to increase the edge-connectivity of a given graph from 1 to 2 in the cheapest possible way by adding some additional edges from a given set. There are many standard techniques that lead to a 2-approximation for WTAP, but despite much progress on special cases, the factor 2 remained unbeaten for several decades. <br> In this talk we present two algorithms for WTAP that improve on this longstanding approximation ratio of 2. The first algorithm is a relative greedy algorithm, which starts with a simple, though weak, solution and iteratively replaces parts of this starting solution by stronger components. <br> This algorithm achieves an approximation ratio of (1+ln(2)+ϵ)<1.7. Second, we present a local search algorithm that achieves an approximation ratio of 1.5+ϵ (for any constant ϵ>0). <br> This is joint work with Rico Zenklusen. <br> Microsoft Teams Link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> For more details about the seminar please visit the website at <a href="https://www.csa.iisc.ac.in/iisc-msr-seminar/">https://www.csa.iisc.ac.in/iisc-msr-seminar/</a> DTSTART:20210910T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210913T120000Z UID:b61e8bcaca2cf12ddc5e341b51ecacbd-196 DTSTAMP:19700101T120011Z DESCRIPTION:GPM: Exploring GPUs with Persistent Memory URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/196/gpm-exploring-gpus-with-persistent-memory/ SUMMARY:GPUs are a key computing platform for many application domains. While the emergence of non-volatile memory has brought the promise of fi ne-grain byte-addressable persistence (a.k.a., persistent memory, or PM) to CPU applications, the same unfortunately is beyond the reach of GPU programs. <br> This work takes three steps toward enabling GPU programs to directly access PM. First, we show how various existing software and hardware technologies can be put together to enable direct access to PM from within a GPU kernel. Next, we created a workload suite of 10 GPU-accelerated applications to demonstrate how GPUs can benefit from PM. We then created a GPU library to support logging, checkpointing, and primitives for native persistence to enable GPU programmers to easily leverage PM. DTSTART:20210913T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210917T120000Z UID:e683356c19913b24273a003a3425018b-197 DTSTAMP:19700101T120016Z DESCRIPTION:Efficient Intervention Design for Learning Causal DAGs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/197/efficient-intervention-design-for-learning-causal-dags/ SUMMARY:Network of cause effect relationships between measured variables is modeled as a causal DAG (Directed Acyclic Graph). In this talk, we focus on efficient adaptive intervention design for learning a causal DAG, with no latent confounders, given the observational equivalence class it belongs to as an input. We first consider equivalence class inputs whose skeleton is a tree. We consider a Bayesian framework where a prior over all directed trees is updated based on the outcome of every single node intervention each of which is adaptively designed based on current posterior. We provide an efficient algorithm that requires interventions that is within a factor of 2 from the best adaptive algorithm. We also show that information greedy approaches are exponentially sub-optimal in terms of the optimal number of interventions required. The main technical tool is a simple greedy algorithm that myopically optimizes a centrality measure on the skeleton of the true causal tree. <br> We generalize and extend the above approach for adaptive interventional design to learn an arbitrary causal DAGs given its equivalence class. We show that the half the maximum clique size is an instance specific fundamental lower bound for any algorithm to even verify the DAG structure through interventions given the equivalence class. Under mild assumptions on the equivalence class, we provide an adaptive algorithm inspired by the algorithm on causal trees that requires interventions that matches the optimal number of interventions up to a multiplicative logarithmic factor in the number of maximal cliques. <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20210917T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210916T120000Z UID:7f89995347e9834c0e1a5b0afd5eb20b-198 DTSTAMP:19700101T120016Z DESCRIPTION:Security of Post-Quantum Multivariate Blind Signature Scheme: Revisited and Improved URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/198/security-of-post-quantum-multivariate-blind-signature-scheme-revisited-and-improved/ SUMMARY:Current ubiquitous cryptosystems face an imminent threat from quantum algorithms like Shors and Grovers, leading us to a future of post-quantum cryptography. Multivariate signatures are prominent in post-quantum cryptography due to their fast, low-cost implementations and shorter signatures. Blind signatures are a special category of digital signatures with two security notions: blindness and one-more unforgeability (OMF). Our work primarily focuses on the multivariate blind signature scheme (MBSS) proposed by Petzoldt et al. We construct a formal proof along the lines of the heuristic sketch given by the authors of MBSS based on the proposed universal one-more unforgeability (UOMF) model, a weaker variant of OMF. The construction of this proof led us to identify the various issues in the security argument - mainly the difficulty in simulating the response to the blind signature queries without knowing the secret key of the underlying Rainbow scheme used. Since our investigation revealed the difficulty in reducing the UOMF security to the hardness assumption used by the authors, therefore we design a new class of hardness assumptions: (1) Single Target Inversion Problem, PR-STI (2) Modified version of Single Target Inversion Problem, PR-mSTI (3) Chosen Target Inversion Problem, PR-CTI. Armed with the new class of computational problems, we reduce the UOMF and OMF security of MBSS to these problems. We begin by giving an improved security argument of MBSS in the UOMF security model using the PR-mSTI assumption. We employ the general forking algorithm in this security reduction. However, we cannot apply the forking algorithm directly owing to the wrapper algorithm being split and the presence of the blind signature oracle. We thus suitably modify the algorithm and then derive the corresponding forking probability. To argue the security of MBSS in the standard security model, i.e., in the OMF model, we try using the PR-CTI assumption. The PR-CTI problem demands computing the solution for more than one target. With the high cost of forking, a significant degradation factor appears in the success probability. So, instead, we reduce the OMF security of MBSS to the PR-mSTI assumption (in a restricted setting) and give a comparative analysis between the security argument in the UOMF and OMF models. <br> <br> Microsoft Teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzFiMjNmODEtN2ZkZi00OGU4LWEzOGEtZTdjMjM5OWI0ZWZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22f24581ee-9350-45bb-b76b-26beecd9d45f%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzFiMjNmODEtN2ZkZi00OGU4LWEzOGEtZTdjMjM5OWI0ZWZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22f24581ee-9350-45bb-b76b-26beecd9d45f%22%7d</a> DTSTART:20210916T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210922T120000Z UID:f648670e9a1485985fe0b30228840b9d-199 DTSTAMP:19700101T120008Z DESCRIPTION:Designing Secure Cryptographic Systems: Journey from Theory to Practice URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/199/designing-secure-cryptographic-systems-journey-from-theory-to-practice/ SUMMARY:The study of cryptography is aimed at keeping information secure in an increasingly digitized world. Modern cryptography uses theoretical frameworks to prove the security of cryptographic primitives against precisely modeled attacks. However, translating cryptographic primitives from provably secure algorithms into secure deployable systems remains a massive challenge. In particular, existing theoretical models do not account for potential weaknesses inherent to practical cryptographic implementations. Hence, provable security guarantees often collapse in the face of attacks that exploit implementation-level weaknesses to devastating effect. <br> In this talk, I will give an overview of my journey so far in attempting to bridge the wonderfully multi-faceted aspects of cryptography, with the aim of designing, analyzing and securely implementing cryptographic solutions to real-world problems while relying on as minimal a set of assumptions as possible. In the process, I will summarize my past research works spanning theoretical cryptographic foundations, applied cryptography and secure cryptographic implementations. <br> I will begin with an overview of my foundational research into enabling a variety of functionally rich and provably secure cryptographic applications based on Minicrypt (the world of “symmetric-key” cryptoprimitives), and some additional algebraic structure. I will then discuss my research efforts towards enabling a specific cryptographic application - searchable symmetric encryption (SSE) - that supports a wide class of Boolean queries over encrypted relational databases at scale while relying on purely symmetric-key primitives. Finally, I will showcase that despite the theoretical security guarantees afforded by standardized symmetric-key cryptographic algorithms such as AES-128, practical implementations of SSE schemes remain vulnerable to "fault-injection attacks – a special class of implementation-level attacks powerful enough to reduce the keyspace for AES-128 from 2^{128} to a single key while relying on a single fault-injection. In particular, I will describe my recent work (appeared at Eurocrypt 2020) on a “fault propagation”-based key-recovery attack that completely breaks the security of an AES-128 implementation, even when equipped with dedicated protections against standard implementation-level attacks. <br> No prior background on cryptography will be needed. DTSTART:20210922T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210923T120000Z UID:1af4a6d706c83cf5eaecff7fefbd20c3-200 DTSTAMP:19700101T120016Z DESCRIPTION:Vertex Coloring Problem with Some Domination Properties URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/200/vertex-coloring-problem-with-some-domination-properties/ SUMMARY:Coloring and Dominating Set are two well known and well studied problems in Graph Theory. Here we consider a vertex coloring problem by imposing some domination property on it. Given a simple graph G = (V,E), in domination coloring, we need to find a vertex coloring of V such that each vertex v in V dominates some color class and each color class is dominated by some vertex v in V. The domination chromatic number of G, is the minimum number of colors used in a domination coloring of G. In minimum domination coloring we need to compute a domination coloring with minimum number of colors. We prove that, this problem is as hard as minimum vertex coloring problem. We also show that, it cannot be approximated within a factor of O(ln n), even when restricted to weakly chordal graphs. We also consider node deletion problems associated with domination coloring. Given a graph G and a positive integer q, in Minimum q-Domination Partization, it is required to find a vertex set S of minimum size such that the remaining graph is domination colorable with at most q colors. We only consider q-domination partization problem for q = 2 and q =3. We prove that Minimum 2-Domination Partization is APX-complete. It is approximable within a factor of 2 and this approximation factor is the best possible approximation factor. Then we have given the characterizations for 3-domination colorable graphs. Finally, we prove that Minimum 3-Domination Partization is APX-hard, it is equivalent to minimum odd cycle transversal and design approximation algorithm for it. DTSTART:20210923T120000Z END:VEVENT BEGIN:VEVENT DTEND:20210924T120000Z UID:13d33582b1e92e8cc4b577237f70e031-201 DTSTAMP:19700101T120016Z DESCRIPTION:Monotone Arithmetic Lower Bounds Via Communication Complexity. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/201/monotone-arithmetic-lower-bounds-via-communication-complexity/ SUMMARY:How much power does negation or cancellation provide to computation? <br> This is a fundamental question in theoretical computer science that <br> appears in various parts: in Boolean circuits, Arithmetic circuits and <br> also in communication complexity. I will talk about some new connections <br> between the latter two fields and their applications to extend two <br> classical results from four decades ago: Valiant (1979) showed that <br> monotone arithmetic circuits are exponentially weaker than general <br> circuits for computing monotone polynomials. Our first result gives a <br> qualitatively more powerful separation by showing an exponential <br> separation between general monotone circuits and constant-depth <br> multi-linear formulas. Neither such a separation between general <br> formulas and monotone circuits, nor a separation between multi-linear <br> circuits and monotone circuits were known before. Our result uses the <br> recent counter-example to the Log-Approximate-Rank Conjecture in <br> communication complexity. <br> <br> <br> Jerrum and Snir (1982) also obtained a separation between the powers of <br> general circuits and monotone ones via a different polynomial, i.e. the <br> spanning tree polynomial (STP), a polynomial that is well known to be in <br> VP, using non-multi-linear cancellations of determinantal computation. <br> We provide the first extension of this result to show that the STP <br> remains `robustly hard` for monotone circuits in the sense of Hrubes <br> recent notion of epsilon-sensitivity. The latter result is proved via <br> formulating a discrepancy method for monotone arithmetic circuits that <br> seems independently interesting. <br> <br> We will discuss several open problems arising from these results. <br> <br> (These are based on joint works with Rajit Datta, Utsab Ghosal and <br> Partha Mukhopadhyay). <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20210924T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211001T120000Z UID:c3193587405e85a5d199b2c2bde43d0f-202 DTSTAMP:19700101T120016Z DESCRIPTION:Improved (exponential time) algorithms: A case study for Subset Sum and Bin Packing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/202/improved-exponential-time-algorithms-a-case-study-for-subset-sum-and-bin-packing/ SUMMARY:Given an algorithm for a computational problem, a natural question is: Can its time or space efficiency be improved? We study this question for some natural and/or old algorithms for NP-complete problems. <br> Specifically, we survey some of the modern techniques to design such improved algorithms, with a focus on the Subset Sum and Bin Packing problems: <br> The algorithm by Schroeppel and Shamir (FOCS'79) solving Subset Sum on instances with n items in O(2^{n/2}) time and O(2^{n/4}) space can be improved to an algorithm using the same time and O(2^{0.249999n}) space. The trivial O(2^n) time and poly(n) space algorithm for Subset Sum can be improved to an O(2^{0.86n}) time poly(n) space algorithm, assuming random read-only access to random bits. The standard algorithm solving Bin Packing with n items in O(2^n) can be improved to an algorithm running in time O((2−ε_b)^n), where n denotes the number of items and ε_b is a positive number that only depends on the number of bins b available in the instance. <br> Two key modern techniques we will discuss are (1) a new method based on anti-concentration of subset sums (along with structural new insights in additive combinatorics) and (2) the representation method by Joux and Howgrave-Graham (EUROCRYPT'10) to navigate through the search space in an improved way, <br> <br> We will discuss parts of the following works: <br> Jesper Nederlof, Jakub Pawlewicz, Céline M. F. Swennenhuis, Karol Wegrzycki: A Faster Exponential Time Algorithm for Bin Packing With a Constant Number of Bins via Additive Combinatorics. SODA 2021: 1682-1701. <br> Jesper Nederlof, Karol Wegrzycki: Improving Schroeppel and Shamir's algorithm for subset sum via orthogonal vectors. STOC 2021: 1670-1683 <br> Nikhil Bansal, Shashwat Garg, Jesper Nederlof, Nikhil Vyas: Faster space-efficient algorithms for subset sum and k-sum. STOC 2017: 198-209 <br> Per Austrin, Petteri Kaski, Mikko Koivisto, Jesper Nederlof: Dense Subset Sum May Be the Hardest. STACS 2016: 13:1-13:14 <br> <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20211001T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211005T120000Z UID:9278b801204eabedb45805f551a26280-203 DTSTAMP:19700101T120009Z DESCRIPTION:A Framework for Privacy-Compliant Delivery Drones URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/203/a-framework-for-privacy-compliant-delivery-drones/ SUMMARY:This thesis presents Privaros, a framework to enforce privacy policies on drones. Privaros is designed for commercial delivery drones, such as the ones that will likely be used by Amazon Prime Air. Such drones visit various host airspaces, each of which may have different privacy requirements. Privaros uses mandatory access control to enforce the policies of these hosts on guest delivery drones. Privaros is tailored for ROS, a middleware popular in many drone platforms. This thesis presents the design and implementation of Privaross policy-enforcement mechanisms, describes how policies are specified, and shows that policy specification can be integrated with Indias Digital Sky portal. This thesis presents an evaluation of Privaros that shows that a drone running Privaros can robustly enforce various privacy policies specified by hosts, and that its core mechanisms only marginally increase communication latency and power consumption. <br> <br> Microsoft Teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MWNhNDA3ODYtNWI3NC00Zjg3LWJjM2YtYzBiNTQwOTI2ZTBk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_MWNhNDA3ODYtNWI3NC00Zjg3LWJjM2YtYzBiNTQwOTI2ZTBk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d</a> DTSTART:20211005T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211008T120000Z UID:d75dd7006f1bf6c4dc9a3a5a2c29559c-204 DTSTAMP:19700101T120010Z DESCRIPTION:A Multi-Policy Reinforcement Learning Framework for Autonomous Navigation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/204/a-multi-policy-reinforcement-learning-framework-for-autonomous-navigation/ SUMMARY:Reinforcement Learning (RL) is the process of training an agent to take a sequence of actions with the prime objective of maximizing rewards it obtains from an environment. Deep RL is simply using the same approach where a deep neural network parameterizes the policy. Temporal abstraction in RL is learning useful and generalizable skills, which are often necessary for solving complex tasks in various environments of practical interest. One such domain is the longstanding problem of autonomous vehicle navigation. In this work, we focus on learning complex skills in such environments where the agent has to learn a high-level policy by leveraging multiple skills inside an environment that presents various challenges. <br> Multi-policy reinforcement learning algorithms like the Options Critic Framework require an exorbitant amount of time for converging to policies. Even when they do, there is a broad tendency for the policy over options to choose a single sub-policy exclusively, thus rendering the other policies moot. In contrast, our approach combines an iterative approach to complement previously learned policies. <br> To conduct the experiments, a custom simulated 3D navigation environment was developed where the agent is a vehicle that has to learn a policy by which it can avoid a collision. This is complicated because, in some scenarios, the agent needs to infer certain abstract meaning from the environment to make sense of it while learning from a reward signal that becomes increasingly sparse. <br> In this thesis, we introduce the `Stay Alive` approach to learn such skills by sequentially adding them into an overall set without using an overarching hierarchical policy where the agents objective is to prolong the episode for as long as possible. The general idea behind our approach comes from the fact that both animals and human beings learn meaningful skills in previously acquired skills to better adapt to their respective environments. <br> We compare and report our results on the navigation environment and the Atari Riverraid environment with state-of-the-art RL algorithms and show that our approach outperforms the prior methods. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NTM0NzExMWMtNmFmYi00OWZjLTliM2EtNTNjYTQyNzg4OTVm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229d60e185-2600-4b28-b2d5-2d47a54928f3%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_NTM0NzExMWMtNmFmYi00OWZjLTliM2EtNTNjYTQyNzg4OTVm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229d60e185-2600-4b28-b2d5-2d47a54928f3%22%7d</a> DTSTART:20211008T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211008T120000Z UID:9f39569e7c6ed35c844d58bc2fc568d0-205 DTSTAMP:19700101T120016Z DESCRIPTION:Algorithmic advances on metric and graph clustering (Part 1) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/205/algorithmic-advances-on-metric-and-graph-clustering-part-1/ SUMMARY:Clustering algorithms are at the core of unsupervised machine learning and data analysis techniques. <br> Given a set of data elements, the goal of a clustering is to partition a dataset in such a way that <br> data elements in the same part are more similar to each other than data elements in different parts. <br> Clustering problems arise in large variety of applications ranging from bioinformatics to computer vision <br> and as such are very basic problems. <br> <br> <br> In these two talks, we will present both metric clustering (Part 1) and graph clustering (Part 2) problems. <br> We will first illustrate some recent advances in the complexity of the classic k-median and k-means problems, <br> two popular objective functions for metric clustering, via some recent developments on the fixed-parameter <br> tractability of the objectives and hardness of approximation. We will then describe new approximation algorithms <br> for metric hierarchical clustering. <br> <br> In the second part of the talks, we will present a new perspective on the classic correlation clustering <br> objective that leads to new efficient distributed algorithms for the problem, together with a beyond-the-worst-case <br> analysis of the Louvain algorithm for finding the maximum modularity graphs clustering. <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d </a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20211008T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211014T120000Z UID:6f9fd9b4a3817fc93b4819a1fc4de560-206 DTSTAMP:19700101T120009Z DESCRIPTION:A Trusted-Hardware Backed Secure Payments Platform for Android URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/206/a-trusted-hardware-backed-secure-payments-platform-for-android/ SUMMARY:Digital payments using personal electronic devices have been steadily gaining in popularity for the last few years. While digital payments using smartphones are very convenient, they are also more susceptible to security vulnerabilities. Unlike devices dedicated to the purpose of payments (e.g. POS terminals), modern smartphones provide a large attack surface due to the presence of so many apps for various use cases and a complex feature-rich smartphone OS. <br> Because it is the most popular smartphone OS by a huge margin, Android is the primary target of attackers. Although the security guarantees provided by the Android platform have improved significantly with each new release, we still see new vulnerabilities being reported ever month. Vulnerabilities in the underlying Linux kernel are particularly dangerous because of their severe impact on app security. To protect against a compromised kernel, some critical functions of the Android platform such as cryptography and local user authentication have been moved to a Trusted Execution Environment (TEE) in the last few releases. But the Android platform doesn't yet provide a way to protect a user's confidential input meant for a remote server, from a compromised kernel. Our work aims to address this gap in Android's use of TEEs for app security. We have designed a framework that leverages a TEE for protecting user's confidential input and we have shown how this framework can be used to improve the security of digital payments, <br> <br> Microsoft teams link: <br><a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTVlNGMzNzAtOWU3MC00ODA5LTk0MTAtNDNhZDc5YjNlMGFm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTVlNGMzNzAtOWU3MC00ODA5LTk0MTAtNDNhZDc5YjNlMGFm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220432f3a0-d225-405c-b0f4-ff1ffaf4f1fd%22%7d</a> DTSTART:20211014T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211018T120000Z UID:4a9818f675917650a27b11744c8319ed-207 DTSTAMP:19700101T120016Z DESCRIPTION:Algorithmic advances on metric and graph clustering (Part 2) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/207/algorithmic-advances-on-metric-and-graph-clustering-part-2/ SUMMARY:Clustering algorithms are at the core of unsupervised machine learning and data analysis techniques. <br> Given a set of data elements, the goal of a clustering is to partition a dataset in such a way that <br> data elements in the same part are more similar to each other than data elements in different parts. <br> Clustering problems arise in large variety of applications ranging from bioinformatics to computer vision <br> and as such are very basic problems. <br> <br> <br> In these two talks, we will present both metric clustering (Part 1) and graph clustering (Part 2) problems. <br> We will first illustrate some recent advances in the complexity of the classic k-median and k-means problems, <br> two popular objective functions for metric clustering, via some recent developments on the fixed-parameter <br> tractability of the objectives and hardness of approximation. We will then describe new approximation algorithms <br> for metric hierarchical clustering. <br> <br> <br> In the second part of the talks, we will present a new perspective on the classic correlation clustering <br> objective that leads to new efficient distributed algorithms for the problem, together with a beyond-the-worst-case <br> analysis of the Louvain algorithm for finding the maximum modularity graphs clustering. <br> <br> Microsoft Teams Link:<br> <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> The first part of this talk is uploaded on CSA YouTube channel: https://www.youtube.com/watch?v=7vKYZFjGwo8 <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20211018T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211022T120000Z UID:b94b802a0001c4e68ea0b8d5a6c21553-208 DTSTAMP:19700101T120016Z DESCRIPTION:Recent developments on ∃R-completeness of packing and other problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/208/recent-developments-on-a%cb%86%c6%92r-completeness-of-packing-and-other-problems/ SUMMARY:We will give an introduction to the complexity class ∃R, which consists of problems that are polynomial time reducible to deciding whether system of polynomial equations and inequalities with integer coefficients and many unknowns has a real solution. Many classic problems have recently been shown to be ∃R-complete, such as the Art Gallery Problem, the Minimum Convex Cover problem, training neural networks, geometric embeddability of simplicial complexes, and many variants of 2D packing problems. We will outline some of the techniques used in these proofs, in particular for the case of the ∃R-hardness of packing problems. <br> <br> <br> Microsoft Teams Link: <br> <a href"https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20211022T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211026T120000Z UID:6ec75329a9208e468835b65939c7b498-209 DTSTAMP:19700101T120011Z DESCRIPTION:Novel Reinforcement Learning Algorithms and Applications to Hybrid Control Design Problems. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/209/novel-reinforcement-learning-algorithms-and-applications-to-hybrid-control-design-problems/ SUMMARY:The thesis is a compilation of two independent works. In the first work, we develop novel framework that allows assignment of different weights to different $n$ step returns, which helps us develop several schedule based algorithms. Learning the value function of a given policy from the data samples is an important problem in Reinforcement Learning.<br> TD($lambda$) is a popular class of algorithms to solve this problem. However, the weight assigned to different $n$-step returns decreases exponentially with increasing $n$ in TD($lambda$). Here, we present a $lambda$-schedule procedure that allows flexibility in weight assignment to the different $n$-step returns. <br> Based on this procedure, we propose an on-policy algorithm, TD($lambda$)-schedule, and an off-policy algorithm, TDC($lambda$)-schedule, respectively. We provide proofs of almost sure convergence for both algorithms under a general Markov noise framework as well as present the results of experiments where these algorithms are seen to show improved performance.<br> In the second work, we design hybrid control policies for hybrid systems whose mathematical models are unknown.<br> <br> Our contributions are threefold here. <br> First, we propose a framework for modelling the hybrid control design problem as a single Markov Decision Process (MDP).<br> This result facilitates the application of off-the-shelf algorithms from Reinforcement Learning (RL) literature towards designing optimal control policies. <br> Second, we model a set of benchmark examples of hybrid control design problem in the proposed MDP framework. <br> Third, we adapt the recently proposed Proximal Policy Optimisation (PPO) algorithm for the hybrid action space and apply it to the above set of problems. It is observed that in each case the algorithm converges and finds the optimal policy. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MmE2OTllMjEtNDZiMS00MzNhLWI1OTYtMGM2MWE5OTY5ZmEw%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%225df91a8a-f4f7-4407-b829-a2fd5472e572%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MmE2OTllMjEtNDZiMS00MzNhLWI1OTYtMGM2MWE5OTY5ZmEw%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%225df91a8a-f4f7-4407-b829-a2fd5472e572%22%7d</a> DTSTART:20211026T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211029T120000Z UID:ab8de16382e7a652aecd905e8e80ff60-210 DTSTAMP:19700101T120011Z DESCRIPTION:Designing CPUs For Leadership High-Performance Computing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/210/designing-cpus-for-leadership-high-performance-computing/ SUMMARY:Microprocessors are playing a central role in this era of exponential growth in computing. CPU designs are becoming increasingly complex as they evolve in response to the demanding performance requirements of existing and emerging applications and deployment models (e.g. cloud services). AMD is at the forefront of the high-performance CPU design and is delivering leadership performance with Ryzen and EPYC series of CPUs for the client and server markets respectively. In this talk, we will provide an overview of AMD’s CPU microarchitecture and discuss some of the problems we are working on to further improve CPU performance. We will also discuss the process and challenges of designing high-performance CPUs for the fast-evolving application and design space, highlighting problems related to microarchitecture, simulation technology, workload sampling, and performance projections. DTSTART:20211029T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211103T120000Z UID:5751a9e273e5ac6bc4af1b91cd0129c6-211 DTSTAMP:19700101T120015Z DESCRIPTION:Statistical Network Analysis: Community Structure, Fairness Constraints, and Emergent Behavior URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/211/statistical-network-analysis-community-structure-fairness-constraints-and-emergent-behavior/ SUMMARY:Networks or graphs provide mathematical tools for describing and analyzing relational data. They are used in biology to model interactions between proteins, in economics to identify trade alliances among countries, in epidemiology to study the spread of diseases, and in computer science to rank webpages on a search engine, to name a few. Each application domain in this wide assortment encounters networks with diverse properties and imposes various constraints. For example, networks may be dynamic, heterogeneous, or attributed, and an application domain may require a fairness constraint on the communities (e.g., requiring communities in a social network to be balanced with respect to genders). However, most existing research is concerned with the simplest type of networks with a fixed set of nodes and edges and focuses on the canonical forms of tasks like community detection and link prediction. This thesis aims at bridging this gap between the simplistic problem settings considered in the literature and the complex requirements of real-world applications by proposing community detection and link prediction methods to analyze different types of networks from various perspectives. <br> Our first contribution includes two spectral algorithms for finding `fair` communities in a given network G. We define what it means for communities to be fair from the perspective of each node (a.k.a. individual fairness). This is done via an auxiliary `representation graph R that connects nodes if they can represent each others interests in various communities. Informally speaking, a node finds communities fair if its neighbors from R are proportionally distributed across all communities in G. The goal is to find communities that are considered fair by all nodes. We show that this fairness criterion (i) generalizes the well-explored idea of statistical fairness and (ii) is also applicable in cases where sensitive node attributes (like gender and race) are not observable but instead manifest themselves as intrinsic or latent features in R. We develop fair spectral clustering algorithms and prove that they are weakly consistent (#mistakes = o(N) with probability 1 - o(1)) under a proposed variant of the stochastic block model. <br> Second, we propose a community-based statistical model for dynamic networks where edges appear and disappear over time. Many networks like social networks, citation networks, contact networks, etc., are dynamic in nature. Our model embeds the nodes and communities in a d-dimensional latent space and specifies a procedure for updating these embeddings over time to model the networks evolution. Given an observed dynamic network, we infer these latent quantities using variational inference and use them for link forecasting and community detection. Unlike existing approaches, our model supports the birth and death of communities. It also allows us to use powerful neural networks during inference. Experiments demonstrate that our model is better at link forecasting and community detection as compared to existing approaches. Moreover, it discovers stable communities, as quantified by the normalized mutual information (NMI) score between communities discovered at successive time steps. This desirable quality is absent in methods that ignore the network dynamics. <br> Third, we propose a statistical model for heterogeneous dynamic networks where the nodes and relations additionally have a type associated with them (e.g., knowledge graphs). Besides the latent node attributes, this model also encodes a set of interaction matrices for each type of relation. These matrices specify the affinity between nodes based on their attribute values and can represent both homophyllic (like attracts like) and heterophyllic relationships (opposites attract). We develop a scalable neural network-based inference procedure for this model and demonstrate that it outperforms existing state-of-the-art approaches on several homogeneous and heterogeneous dynamic network datasets, particularly the temporal knowledge graphs. <br> Fourth, we develop a model for networks with node covariates to bring explainability to community detection. This model integrates node covariates into a stochastic block model using restricted Boltzmann machines. We subscribe to the view that a community can be explained by identifying the defining covariates of its member nodes. Our model provides the relative importance of various covariates in each community, thereby explaining its decision to group the members. Existing approaches for modeling networks with covariates lack this property, especially the ones that are based on deep neural networks. We also derive an efficient inference procedure that runs in linear time in the number of nodes and edges. Experiments confirm that our models community detection performance is comparable with recent deep neural network-based approaches. However, it additionally offers the advantage of explainability. <br> The discussion till this point views communities as passive structures arising out of interactions between nodes. However, just like existing links in a network determine future links, communities also play a functional role in shaping the behavior of the nodes (for example, preference for a clothing brand). Our final contribution explores this functional view of communities and shows that they affect emergent communication in a networked multi-agent reinforcement learning setting. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MTZhMmY1OTYtNDNhYS00NzM0LWE1ZTYtNTlhODAzNjg1MTQ3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c747ccaa-ceaa-4197-b4cb-ce2f1d4694da%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MTZhMmY1OTYtNDNhYS00NzM0LWE1ZTYtNTlhODAzNjg1MTQ3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c747ccaa-ceaa-4197-b4cb-ce2f1d4694da%22%7d</a> DTSTART:20211103T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211110T120000Z UID:dbdc6cf203514866672ff32ea3f5f409-212 DTSTAMP:19700101T120008Z DESCRIPTION:Towards Model Understanding URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/212/towards-model-understanding/ SUMMARY:While deep learning models have become increasingly accurate over the last decade, concerns about their (lack of) interpretability have taken a center stage. In response, a growing subfield on interpretability and analysis of these models has emerged. Interpretability is an umbrella term encompassing efforts to understand the learned models and communicate that understanding to the stakeholders. In this talk, I will share our research towards these goals and first highlight methods that aid user understanding, and then, focus on protocols to evaluate model explanations—a fundamental issue facing much of interpretability research. <br> Online Teams Meeting Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDljNWVlMTYtYjIzYy00MGM0LWE2YzgtMjBmMDNiYjhhMzJm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDljNWVlMTYtYjIzYy00MGM0LWE2YzgtMjBmMDNiYjhhMzJm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20211110T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211115T120000Z UID:2cb31add9e8ab177527184d58d91dbb2-213 DTSTAMP:19700101T120014Z DESCRIPTION:A Syntactic Neural Model for Question Decomposition URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/213/a-syntactic-neural-model-for-question-decomposition/ SUMMARY:Question decomposition along with single-hop Question Answering (QA) system serve as useful modules in developing multi-hop Question Answering systems, mainly because the resulting QA system is interpretable and has been demonstrated to exhibit better performance. The problem of Question Decomposition can be posed as a machine translation problem, and it can be solved using any sequence-to-sequence neural architecture. Using this approach, it is difficult to capture the innate hierarchical structure of the decomposition. Inspired by database query languages a pseudo-formalism for capturing the meaning of questions, called Question Decomposition Meaning Representation (QDMR) was recently introduced. In this approach, a complex question is decomposed into simple queries which are mapped into a small set of formal operations. This method does not utilize the underlying syntax information of QDMR to generate the decomposition. <br> In the area of programming language code generation, methods that use syntax information as a prior knowledge have been demonstrated to perform better. Moreover, the syntax-aware models are usually interpretable. Motivated by the success of syntax-aware models, we propose a new syntactic neural model for question decomposition in this thesis. In particular, we encode the underlying syntax of the QDMR structures into a grammar model as a sequence of actions. This is done using a deterministic framework which uses Abstract Syntax Trees (AST) and Parse Trees. The proposed approach can be thought of as an encoder-decoder method for QDMR structures where a sequence of possible actions is a latent representation of the QDMR structure. The advantage of using this latent representation is that it is interpretable. Experimental results on a real-world dataset demonstrate that the proposed approach outperforms the state-of-the-art approach especially in scenarios where training data is limited. Some heuristics to further improve the performance of the proposed approach are also suggested in this work. <br> <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_M2ZlNGMxM2YtYTc1NS00ZDQ1LWE5YTktMmY3NTg0ZDg3NDc3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_M2ZlNGMxM2YtYTc1NS00ZDQ1LWE5YTktMmY3NTg0ZDg3NDc3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20211115T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211112T120000Z UID:b97af34e7df4883b5d465339e8e77d30-214 DTSTAMP:19700101T120016Z DESCRIPTION:Min Sum Set Cover and Tail Bounds for Sums of Bernoullis URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/214/min-sum-set-cover-and-tail-bounds-for-sums-of-bernoullis/ SUMMARY:We study the generalized min sum set cover (GMSSC) problem, wherein given a collection of hyperedges E with arbitrary covering requirements, the goal is to find an ordering of the vertices to minimize the total cover time of the hyperedges. We give a 4.642 approximation algorithm for GMSSC, coming close to the best possible bound of 4, improving the previous best bound of 12 by Im, Sviridenko and van der Zwaan. For the special case when the hypergraph is a graph, we give a 16/9 approximation, improving the previous best bound of 1.999946 by Barenholz and Feige. Our algorithms are based on applying a suitable linear transformation on the LP solution and applying randomized rounding. <br> As part of the analysis of our algorithm, we also derive an inequality on the lower tail of a sum of independent Bernoulli random variables, which might be of independent interest and broader utility. Specifically, we show how to get better tail bounds than Chernoff when the deviation from the mean is very small. <br> We also give a new dual-fitting analysis for min sum set cover, giving tight (upto NP-hardness) bound of (p+1)1+1/p(p+1)1+1/p for the â„“pâ„“p norm of cover times. <br> <br> <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20211112T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211126T120000Z UID:7eafd224f7eff509fbbf4fa0086fab27-215 DTSTAMP:19700101T120016Z DESCRIPTION:Performance Characterization and Optimizations of Traditional ML Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/215/performance-characterization-and-optimizations-of-traditional-ml-applications/ SUMMARY:In recent years, Deep Learning based methods have attracted a lot of attention and research – both from statistics and systems. These traditional algorithms are easily explainable and are pretty fast for smaller and medium-size datasets. However, in large organizations, massive datasets spanning a couple of million sample points are not rare. A lot of research has been done to design or adapt these traditional algorithms for such massive datasets. However, we find an apparent lack of a detailed systems-based study for these algorithms in the context of huge datasets. <br> In this work, we study the systems behavior and bottlenecks for these algorithms in the context of huge training datasets. As part of our work, we start with a performance characterization study, identify critical performance bottlenecks experienced by these applications, and then measure the reduction in performance stalls along with apparent benefits in terms of speedup after applying some of the well-known optimizations at the levels of caches, main memory, and computation. More specifically, we apply optimizations such as (i) software prefetching to improve cache performance and (ii) data layout and computation reordering optimizations to improve locality in DRAM accesses and show the performance benefits they can bring in these applications. Last, we evaluate the sensitivity of predictions and the improvement in performance when the computations on precise (float/double) inner variables are interpreted as relatively low-cost integer operations. These optimizations are implemented as modification on the well-known scikit-learn library. <br> We evaluate the impact of the proposed optimizations using a combination of simulation and execution on real system and performance measurement. Our optimizations result in performance benefits varying from 5% -- 27% on different ML applications. <br> This is an online colloquium. The teams meeting link for this is: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NjkwMTc1MzAtYTUzYi00ZTcwLTk0OTQtYmNiOWMxODc4YjRi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NjkwMTc1MzAtYTUzYi00ZTcwLTk0OTQtYmNiOWMxODc4YjRi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20211126T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211124T120000Z UID:0261c72e8e95da2bf33a6ccde8cb8abd-216 DTSTAMP:19700101T120011Z DESCRIPTION:Enhancing Coverage and Robustness of Database Generators URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/216/enhancing-coverage-and-robustness-of-database-generators/ SUMMARY:Generating synthetic databases that capture essential data characteristics of client databases is a common requirement for enterprise database vendors. This need stems from a variety of use-cases, such as application testing and assessing performance impacts of planned engine upgrades. A rich body of literature exists in this area, spanning from the early techniques that simply generated data ab-initio to the contemporary ones that use a predefined client query workload to guide the data generation. In the latter category, the aim specifically is to ensure volumetric similarity -- that is, assuming a common choice of query execution plans at the client and vendor sites, the output row cardinalities of individual operators in these plans are similar in the original and synthetic databases. <br> Hydra is a recently proposed data regeneration framework that provides volumetric similarity. In addition, it also provides a mechanism to generate data dynamically during query execution, using a minuscule database summary. Notwithstanding its desirable characteristics, Hydra has the following critical limitations: (a) limited scope of SQL operators in the input query workload, (b) poor scalability with respect to the number of queries in the input workload, and (c) poor volumetric similarity on unseen queries. The data generation algorithm internally uses a linear programming (LP) solver that throttles the workload scalability. This not only puts a threshold on the training (seen) workload size but also reduces the accuracy for test (unseen) queries. Robustness towards test queries is further adversely affected by design choices such as a lack of preference among candidate synthetic databases, and artificial skew in the generated data. <br> In this work, we present an enhanced version of Hydra, called High-Fidelity Hydra (HF-Hydra), which attempts to address the above limitations. To start with, we expand the SQL operator coverage to also include the LIKE operator, and, in certain restricted settings, projection-based operators such as GROUP BY and DISTINCT. To sidestep the challenge of workload scalability, HF-Hydra outputs not one, but a suite of database summaries such that they collectively cover the entire input workload. The division of the workload into the associated sub-workloads is governed by heuristics that aim to balance robustness with LP solvability. <br> For generating richer database summaries, HF-Hydra additionally exploits metadata statistics maintained by the database engine. Further, the database query optimizer is leveraged to make the choice among the various candidate databases. The data generation is also augmented to provide greater diversity in the represented values. Finally, when a test query is fired, HF-Hydra directs it to the database summary that is expected to provide the highest volumetric similarity. <br> We have experimentally evaluated HF-Hydra on a customized set of queries based on the TPC-DS decision-support benchmark framework. We first evaluated the specialized case where each training query has its own summary, and here HF-Hydra achieves perfect volumetric similarity. Further, each summary construction took just under a second and the summary sizes were just in the order of a few tens of kilobytes. Also, our dynamic generation technique produced gigabytes of data in just a few seconds.<br> For the general setting of a limited set of summaries representing the training query workload, the data generated by HF-Hydra was compared with that from Hydra. We observed that HF-Hydra delivers more than forty percent better accuracy for outputs from filter nodes in the plans, while also achieving an improvement of about twenty percent with regard to join nodes. Further, the degradation in volumetric similarity is minor as compared to the one-summary scenario, while the summary production is significantly more efficient due to reduced overheads on the LP solver. <br> In summary, HF-Hydra takes a substantive step forward with regard to creating expressive, robust, and scalable data regeneration frameworks with immediate relevance to testing deployments. <br> Microsoft Teams Meeting Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MTIyMjViNDItOThmNi00MmMxLThhMWUtODUwOWNhZTRjY2E4%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a4bde2-27ad-4a57-9b14-bb2f66dfd6d0%22%7d"> https://teams.microsoft.com/l/meetup-join/19%3ameeting_MTIyMjViNDItOThmNi00MmMxLThhMWUtODUwOWNhZTRjY2E4%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a4bde2-27ad-4a57-9b14-bb2f66dfd6d0%22%7d</a> DTSTART:20211124T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211129T120000Z UID:fd656801f17e40da90fe9e83d2590739-217 DTSTAMP:19700101T120015Z DESCRIPTION:High-Performance GPU Tensor Core Code Generation for Matmul using MLIR URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/217/high-performance-gpu-tensor-core-code-generation-for-matmul-using-mlir/ SUMMARY:State of the art in high-performance deep learning is primarily driven by highly tuned libraries. These libraries are often hand-optimized and tuned by expert programmers using low-level abstractions with significant effort. A lot of the effort may have to be repeated for similar hard- ware and future ones. This process is thus not modular or reusable to the same extent that compiler infrastructures like LLVM are. Manual optimization does not typically use a standard intermediate representation (IR) or transformations and passes on such IRs, although the optimizations performed can be encoded as a sequence of transformation steps and customized passes on an IR. <br> We believe that until the recent introduction of MLIR (Multi-level intermediate representation), IR infrastructure was not geared to tackle the problem of automatic generation of libraries in an effective manner. In particular, it was hard to represent and transform compute abstractions at high, middle, and low levels using a single IR. Multiple levels of abstractions in a single IR permits the user to apply transformations and optimizations at the most suitable level and even reuse them for different targets or front-ends. <br> Some previous works have optimized matrix-matrix multiplication (matmul) for different GPU microarchitectures. All of these works exploit really low-level details of the hardware. Some of them are written directly in assembly, while some use a combination of CUDA C++ with inline assembly. While the set of high-level optimizations is the same, the very dependence on low-level hardware details drifts them away from re-usability. Going against this trend, we show that, by using a set of simple optimizations, suitable abstractions, and lowering passes on such abstractions in MLIR, we can get competitive performance with hand-written libraries. <br> To achieve this, we put together a lowering pipeline that can automatically generate (with- out hand-writing any code) code for matmul on NVIDIA GPUs while utilizing its tensor cores. We have used and extended some existing utilities in MLIR, such as tiling, loop unrolling, loop permutation, and generation of fast memory buffers for input operands. Additional utilities, types, and operations necessary for optimal code generation were implemented from scratch. These include adding WMMA operations and types to provide fundamental support for programming tensor cores, adding loop normalization support, adding multi-level tiling support in affine dialect, creating WMMA operations to load, compute, and store matrix products in a given matmul nest, detection, and hoisting of invariant WMMA load-store pairs, hiding latency of global to shared data movement, and adding support for mapping and converting parallel loops to warps. <br> On a set of problem sizes we evaluated, performance results show that we can attain performance that is 95-119% and 80-160% of cuBLAS, for FP32 and FP16 accumulate respectively, on NVIDIAs Ampere microarchitecture based GeForce 3090 RTX. A similar evaluation on NVIDIAs Turing-based RTX 2080 Ti revealed that we achieve 86-111% and 72-89% of cuBLAS for FP32 and FP16 accumulate, respectively. <br> We take our approach further by fusing common pointwise operations with matrix-matrix multiplication. This is the first work to demonstrate fusion of operations for tensor core matmul using a systematic IR based approach. Fusion is done with the support of additional WMMA operations, which perform warp level matrix operations such as ReLU and constant addition. We see significant gains on small to medium problem sizes when evaluating our fused kernels against a combination of library kernels and custom kernels. On Ampere, consumer fusion performance ranges from 95% to 167% compared with the respective implementations. Similar ranges on Turing are 84% to 150%. We also present preliminary results, which serve as a proof of concept, for producer fusion, i.e., fusion of pointwise operations on the inputs with matmul. Performance of ReLU on C input fused with matmul against a custom ReLU kernel followed by cuBLAS matmul, ranges from 98% to 138% on Ampere and 91% to 133% on Turing. <br> We believe that these results could be used as a foundation and motivation for further research and development on automatic code and library generation using IR infrastructure for similar specialized accelerators. <br> Microsoft teams online link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTcwZjA3NjktMDA2MC00YjE4LWIwYjAtNzQzNmIzNWQ1OTNl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22171d9abc-cf43-429a-9680-c05b9523fa9a%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTcwZjA3NjktMDA2MC00YjE4LWIwYjAtNzQzNmIzNWQ1OTNl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22171d9abc-cf43-429a-9680-c05b9523fa9a%22%7d</a> DTSTART:20211129T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211130T120000Z UID:1e31561b7624b576c1eb4840e9a4bb45-218 DTSTAMP:19700101T120016Z DESCRIPTION:2-Level Page Tables (2-LPT): A building block for efficient address translation in virtualized environment URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/218/2-level-page-tables-2-lpt-a-building-block-for-efficient-address-translation-in-virtualized-environment/ SUMMARY:Efficient address translation mechanisms are gaining more and more attention as the virtual address range of the processors keeps expanding and the demand for machine virtualization increases with cloud and data center-based services. Traditional radix-tree based address translations can incur significant overheads in big data applications, particularly under virtualization, due to multi-level tree walks and nested translation. The overheads stem primarily from unnecessary generality --- ability to support several hundreds of thousands of virtual memory regions in the virtual address space --- supported by current processors. <br> We observe that in the common case, however, a process's virtual address space contains only a few contiguously allocated sections, which can be efficiently translated using a shallow tree with two levels. We propose such a compact structure, called 2-Level Page Table(2-LPT), which is a key building block for address translation in virtualized environment. A key advantage of 2-LPT is that it maintains two levels of page tables irrespective of the size of the virtual address space. Translating a virtual address (VA) using 2-LPT is fast. A walk on a 2-LPT requires up to two memory accesses. In practice, however, the root level table is well cached in the PWC, thus, single memory access is sufficient. Under native execution, 2-LPT reduces the time spent in page walks by up to 20.9% (9.38% on average) and improves performance by up to 10.1% (1.66% on average) over the conventional four-level radix tree page tables, on a set of memory-intensive applications. <br> 2-LPT is more beneficial under virtualization. The proposed 2-LPT design reduces the cost of nested page walk from 24 to 8 memory accesses. To achieve further reduction, we propose two optimizations: (i) Enhanced Partial Shadow Paging (ePSP) which employs a limited form of shadow paging for the root-level of 2-LPT, and (ii) Host PTE Mirroring (HPM) which allows accessing the host page table entry without performing host page table walk. These allow us to largely avoid slow VM exits while effectively reducing the number of memory access on a nested address translation to just one, on average. 2-LPT speeds up applications by 5.6%-50.9% (24.4%, on average) over the baseline with conventional nested page walks. Importantly, it reduces page walk cycles and execution time of the best performing state-of-the-art proposal by 17.1%-57.1% and by 3.9%-43.9%, respectively. <br> MS Team's Meeting Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZTEwNmU0NmEtYTIxMS00YWFkLTk1ZTktZDY4MGJmMjRhMzZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZTEwNmU0NmEtYTIxMS00YWFkLTk1ZTktZDY4MGJmMjRhMzZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20211130T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211201T120000Z UID:f25e36efcf8355ecf3278fe3128abbe5-219 DTSTAMP:19700101T120016Z DESCRIPTION:Quantum-Safe Identity-Based Signature Scheme in Multivariate Quadratic Setting URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/219/quantum-safe-identity-based-signature-scheme-in-multivariate-quadratic-setting/ SUMMARY:Cryptographic techniques are essential for the security of communication in modern society. Today, nearly all public key cryptographic schemes used in practice are based on the two problems of factoring large integers and solving discrete logarithms. However, as the world grapples with the possibility of widespread quantum computing, these schemes are the ones most threatened. Multivariate Public Key Cryptography is one of the possible candidates for security in a post-quantum society, especially in the area of digital signature. This thesis uses the setting of multivariate cryptography to propose an identity-based signature scheme. <br> <br> Our proposal is based on the Rainbow signature scheme and the multivariate 3-pass identification scheme, both of which have been subjected to scrutiny by cryptographers all over the world and have emerged as strong post-quantum candidates. In our construction, we use the identity of users to generate their private key using Rainbow signature scheme. Thereafter, we use these user private keys to sign messages by applying Fiat-Shamir transform to the 3-pass identification scheme. We support the proposed scheme with suitable proof under appropriate computational assumptions, using the standard notions of security. We study the known attacks against multivariate schemes in general, and Rainbow and MQDSS in particular. We then use this analysis to propose concrete parameter sets for our construction. We implement our proposed scheme on an x86-64 PC platform and provide timing results. Our implementation shows that our construction is both practical and efficient. Thus, our proposed scheme stands as a potential post-quantum multivariate signature candidate in the identity-based setting. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjQzMmQ1N2YtMDY5NC00YTk5LWE1ZGQtOWZjNzFjOTkyYzVi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227b997a2d-ed18-48f7-99c1-eaec40b37793%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjQzMmQ1N2YtMDY5NC00YTk5LWE1ZGQtOWZjNzFjOTkyYzVi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227b997a2d-ed18-48f7-99c1-eaec40b37793%22%7d</a> DTSTART:20211201T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211129T120000Z UID:80622a381441ff6934a0ad1e85df4836-221 DTSTAMP:19700101T120016Z DESCRIPTION:Multi-agent Natural Actor-Critic Reinforcement Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/221/multi-agent-natural-actor-critic-reinforcement-learning/ SUMMARY:Both single-agent and multi-agent actor-critic algorithms are an important class of Reinforcement Learning algorithms. In this work, we propose three fully decentralized multi-agent natural actor-critic (MAN) algorithms. The agents objective is to collectively learn a joint policy that maximizes the sum of averaged long-term returns of these agents. In the absence of a central controller, agents communicate the information to their neighbors via a time-varying communication network while preserving privacy. These are decentralized algorithms as each agent picks actions using local reward information and limited information from other agents. We show the convergence of these algorithms using stochastic approximations approach; these algorithms use linear function approximations. We use the Fisher information matrix to obtain the natural gradients. The Fisher information matrix captures the curvature of the Kullback-Leibler (KL) divergence between polices at successive iterates. We also show that the gradient of this KL divergence between policies of successive iterates is proportional to the objective functions gradient. Our MAN algorithms indeed use this representation of the objective functions gradient. Under certain conditions on the Fisher information matrix, we prove that at each iterate, the optimal value via MAN algorithms can be better than that of the multi-agent actor-critic (MAAC) algorithm using the standard gradients. To validate the usefulness of our proposed algorithms, we present extensive computational experiments. First, we implement all the 3 MAN algorithms on a bi-lane traffic network to reduce the average network congestion. We observe an almost 25% reduction in the average congestion in 2 MAN algorithms; the average congestion in another MAN algorithm is on par with the MAAC algorithm. We also consider a generic 15 agent MARL; the performance of the MAN algorithms is again as good as the MAAC algorithm. We attribute the better performance of the MAN algorithms to their use of the above representation of the objective function. <br> Event will be held online on Microsoft Teams <a href="https://teams.microsoft.com/l/meetup-join/19%3ae7d9a7fa7e3f41478d03d76eb63af61b%40thread.tacv2/1626097067191?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22adc1e56f-56ee-4d24-873f-341c97ae782a%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ae7d9a7fa7e3f41478d03d76eb63af61b%40thread.tacv2/1626097067191?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22adc1e56f-56ee-4d24-873f-341c97ae782a%22%7d</a> DTSTART:20211129T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211203T120000Z UID:7f2b9f41ead02d2bbb21ae2a26912bdb-222 DTSTAMP:19700101T120016Z DESCRIPTION:Some Recent Advances in Dynamic Algorithms for Maximum Matching and Minimum Set Cover (Part 1) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/222/some-recent-advances-in-dynamic-algorithms-for-maximum-matching-and-minimum-set-cover-part-1/ SUMMARY:Consider a dynamic graph G = (V, E) that is undergoing a sequence of edge insertions/deletions. We want to design an algorithm that maintains an (approximately) maximum matching in this dynamic graph G with small update time. Here, the update time of an algorithm refers to the time it takes to handle the insertion/deletion of an edge in G. We will like to ensure that the update time of our algorithm is polylogarithmic in the number of nodes G. <br> This problem has received considerable attention in the past decade. In these two talks, I will present an overview of some recent advances on this question. Specifically, I will describe a simple primal-dual algorithm that maintains a (2+epsilon)-approximate fractional matching in G in polylogarithmic update time (Part 1), and show how to efficiently round this fractional matching into an integral one in the dynamic setting (Part 2). <br> Along the way, I will explain some immediate connections between dynamic fractional matching algorithms and the literature on dynamic set cover. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20211203T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211129T120000Z UID:a34bc696360e33678a420d4e509c4945-223 DTSTAMP:19700101T120016Z DESCRIPTION:Optimal Resource Allocation Problems in Large-Scale Networks URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/223/optimal-resource-allocation-problems-in-large-scale-networks/ SUMMARY:Optimal resource allocation in networks gives rise to some of the most fundamental problems at the intersection of algorithms, stochastic processes, and learning. In this talk, we will discuss our recent contributions to three canonical resource allocation problems, namely caching, routing, and scheduling. First, we will consider the optimal caching problem in both stand-alone and networked settings. However, instead of minimizing the competitive ratio - the classical metric of choice for caching problems, we will look at the problem from an online learning perspective that aims to minimize the regret. We will show that this viewpoint leads to an entirely new class of caching policies with provably better performance than the classical ones. We will also discuss some tight regret lower bounds for this problem. Next, we will consider the problem of throughput-optimal dynamic routing of a broad traffic class, including unicast, multicast, broadcast, and anycast flows on a network with arbitrary link scheduling constraints. We will present a unified algorithmic framework based on precedence relaxations, leading to an efficient policy that provably outperforms the state-of-the-art Backpressure routing algorithm. Finally, we will discuss the user scheduling problem for reliable and fresh information delivery over unreliable wireless channels. However, contrary to the existing literature, which predominantly considers stochastic channels, we investigate a non-stationary environment modeled using a new adversarial framework. We will describe competitive algorithms in this setting along with some tight lower bounds. We will supplement each part of the talk with a set of open problems. DTSTART:20211129T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211210T120000Z UID:06005e41efbe918626d94b9b8b2078ae-224 DTSTAMP:19700101T120016Z DESCRIPTION:Some Recent Advances in Dynamic Algorithms for Maximum Matching and Minimum Set Cover (Part 2) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/224/some-recent-advances-in-dynamic-algorithms-for-maximum-matching-and-minimum-set-cover-part-2/ SUMMARY:Consider a dynamic graph G = (V, E) that is undergoing a sequence of edge insertions/deletions. We want to design an algorithm that maintains an (approximately) maximum matching in this dynamic graph G with small update time. Here, the update time of an algorithm refers to the time it takes to handle the insertion/deletion of an edge in G. We will like to ensure that the update time of our algorithm is polylogarithmic in the number of nodes G. <br> This problem has received considerable attention in the past decade. In these two talks, I will present an overview of some recent advances on this question. Specifically, I will describe a simple primal-dual algorithm that maintains a (2+epsilon)-approximate "fractional" matching in G in polylogarithmic update time (Part 1), and show how to efficiently "round" this fractional matching into an integral one in the dynamic setting (Part 2). <br> Along the way, I will explain some immediate connections between dynamic fractional matching algorithms and the literature on dynamic set cover. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20211210T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211216T120000Z UID:206c3b5a36d07fcefe03e3215fdc76f9-225 DTSTAMP:19700101T120008Z DESCRIPTION:Provable and Efficient Algorithms for Federated, Reinforcement and Batch Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/225/provable-and-efficient-algorithms-for-federated-reinforcement-and-batch-learning/ SUMMARY:In this talk, I will propose and analyze iterative algorithms that are computationally efficient, statistically sound and adaptive to the problem complexity. Three different frameworks in which data is presented to the learner are considered : (a) Federated (Distributed) Learning (FL) setup, where data is only available at the edge, and a center machine learns various models via iteratively interacting with the edge nodes; (b) Reinforcement Learning (RL) framework, where agents successively interacts with the environment to learn an optimal policy and (c) Supervised batch setup, where all the data and label pairs are available to the learner at the beginning. In particular, I will explain my contribution in the FL and RL framework and (very) briefly mention the supervised batch setup. <br> In the Federated Learning (FL) framework, I will address the canonical problems of device heterogeneity, communication bottleneck and adversarial robustness for large scale high dimensional problems, and propose efficient and provable algorithms that obtain the optimal statistical rate. In The Reinforcement Learning (RL) setup, I will focus on the problem of structure adaptation (model selection) for large scale problems---an aspect most practical RL based systems like autonomous cars, recommendation systems crucially require. The model selection problem is studied both with function approximation (parameterized RL) as well as for the generic model based RL setup, and the algorithms we propose obtain (near) optimal regret guarantees. <br> Teams Meeting Link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MGUwNjY3NmEtZWZkYy00NzJkLWIzZjMtOWUzMTAyOTgwNzg3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MGUwNjY3NmEtZWZkYy00NzJkLWIzZjMtOWUzMTAyOTgwNzg3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20211216T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211213T120000Z UID:b23d0874f9a26cc35528cf7117a4c1af-226 DTSTAMP:19700101T120009Z DESCRIPTION:Modeling and verification of database-accessing applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/226/modeling-and-verification-of-database-accessing-applications/ SUMMARY:Databases are central to the functioning of most IT-enabled processes and services. In many domains, databases are accessed and updated via applications written in general-purpose languages, as such applications need to contain the business logic and workflows that are key to the organization. Therefore, automated tools are required not only for creation and testing of database schemas and queries, etc., but also for analysis, testing, and verification of database-accessing applications. In this work we describe a novel approach for modeling, analysis and verification of database-accessing applications. We target applications that use Object Relational Mapping (ORM), which is the common database-access paradigm in most Model- View Controller (MVC) based application development frameworks. In contrast with other approaches that try to directly analyze and prove properties of complex database accessing ORM-based code, our approach infers a relational algebra specification of each controller in the application. This specification can then be fed into any off-the-shelf relational algebra solver to check properties (or assertions) given by a developer. We have implemented this approach as a tool that works for Spring based MVC applications. A preliminary evaluation reveals that the approach is scalable and quite precise. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MDdkODcwY2MtNzM3OS00NWVlLWI1YTItNzVhZTIyMTEyYjZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22cd42250e-1d66-4966-a431-6a8d7d5235ba%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MDdkODcwY2MtNzM3OS00NWVlLWI1YTItNzVhZTIyMTEyYjZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22cd42250e-1d66-4966-a431-6a8d7d5235ba%22%7d</a> DTSTART:20211213T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211220T120000Z UID:c33afea382a6e8cb3adb74f39f43bb2f-227 DTSTAMP:19700101T120008Z DESCRIPTION:Energy-efficient Communication Architectures for beyond von-Neumann DNN Accelerators URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/227/energy-efficient-communication-architectures-for-beyond-von-neumann-dnn-accelerators/ SUMMARY:Data communication plays a significant role in overall performance for hardware accelerators of Deep Neural Networks (DNNs). For example, crossbar-based in-memory computing significantly increases on-chip communication volume since the weights and activations are on-chip. State-of-the-art interconnect methodologies for in-memory computing deploy a bus-based network or mesh-based NoC. Our experiments show that up to 90% of the total inference latency of a DNN hardware is spent on on-chip communication when the bus-based network is used. To reduce communication latency, we propose a methodology to generate an NoC architecture and a scheduling technique customized for different DNNs. We prove mathematically that the developed NoC architecture and corresponding schedules achieve the minimum possible communication latency for a given DNN. Experimental evaluations on a wide range of DNNs show that the proposed NoC architecture enables 20%-80% reduction in communication latency with respect to state-of-the-art interconnect solutions. <br> Graph convolutional networks (GCNs) have shown remarkable learning capabilities when processing data in the form of graph which is found inherently in many application areas. To take advantage of the relations captured by the underlying graphs, GCNs distribute the outputs of neural networks embedded in each vertex over multiple iterations. Consequently, they incur a significant amount of computation and irregular communication overheads, which call for GCN-specific hardware accelerators. We propose a communication-aware in-memory computing architecture (COIN) for GCN hardware acceleration. Besides accelerating the computation using custom compute elements (CE) and in-memory computing, COIN aims at minimizing the intra- and inter-CE communication in GCN operations to optimize the performance and energy efficiency. Experimental evaluations with various datasets show up to 174x improvement in energy-delay product with respect to Nvidia Quadro RTX 8000 and edge GPUs for the same data precision. <br> MS Teams Meeting Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZTIxYTBlY2ItZDkyZS00YjVhLTg0YTEtMzNiN2M3OTA4NjMy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZTIxYTBlY2ItZDkyZS00YjVhLTg0YTEtMzNiN2M3OTA4NjMy%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20211220T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211220T120000Z UID:a26bfaa7f2062851bae70fc5c6772610-229 DTSTAMP:19700101T120016Z DESCRIPTION:Security of Post-Quantum Multivariate Blind Signature Scheme: Revisited and Improved URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/229/security-of-post-quantum-multivariate-blind-signature-scheme-revisited-and-improved/ SUMMARY:Current ubiquitous cryptosystems face an imminent threat from quantum algorithms like Shors and Grovers, leading us to a future of post-quantum cryptography. Multivariate signatures are prominent in post-quantum cryptography due to their fast, low-cost implementations and shorter signatures. Blind signatures are a special category of digital signatures with two security notions: blindness and one-more unforgeability (OMF). Our work primarily focuses on the multivariate blind signature scheme (MBSS) proposed by Petzoldt et al. We construct a formal proof along the lines of the heuristic sketch given by the authors of MBSS based on the proposed universal one-more unforgeability (UOMF) model, a weaker variant of OMF. The construction of this proof led us to identify the various issues in the security argument - mainly the difficulty in simulating the response to the blind signature queries without knowing the secret key of the underlying Rainbow scheme used. Since our investigation revealed the difficulty in reducing the UOMF security to the hardness assumption used by the authors, therefore we design a new class of hardness assumptions: (1) Single Target Inversion Problem, PR-STI (2) Modified version of Single Target Inversion Problem, PR-mSTI (3) Chosen Target Inversion Problem, PR-CTI. Armed with the new class of computational problems, we reduce the UOMF and OMF security of MBSS to these problems. We begin by giving an improved security argument of MBSS in the UOMF security model using the PR-mSTI assumption. We employ the general forking algorithm in this security reduction. However, we cannot apply the forking algorithm directly owing to the wrapper algorithm being split and the presence of the blind signature oracle. We thus suitably modify the algorithm and then derive the corresponding forking probability. To argue the security of MBSS in the standard security model, i.e., in the OMF model, we try using the PR-CTI assumption. The PR-CTI problem demands computing the solution for more than one target. With the high cost of forking, a significant degradation factor appears in the success probability. So, instead, we reduce the OMF security of MBSS to the PR-mSTI assumption (in a restricted setting) and give a comparative analysis between the security argument in the UOMF and OMF models. <br> <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YWM1NzljNGYtOTM1MS00NDY1LTliMDktYTRiZTZlOGU5MWJj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22f24581ee-9350-45bb-b76b-26beecd9d45f%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YWM1NzljNGYtOTM1MS00NDY1LTliMDktYTRiZTZlOGU5MWJj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22f24581ee-9350-45bb-b76b-26beecd9d45f%22%7d</a> DTSTART:20211220T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211214T120000Z UID:9f0cafa0c69fa4e972b64a26c2fd37b1-230 DTSTAMP:19700101T120009Z DESCRIPTION:Deep Learning over Hypergraphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/230/deep-learning-over-hypergraphs/ SUMMARY:Though graphs have been extensively used for modelling real-world relational datasets, they are restricted to pairwise relationships, i.e., each edge connects exactly two vertices. Many real-world relational datasets such as academic networks, chemical reaction networks, email communication networks contain group-wise relationships that go beyond pairwise associations. Hypergraphs can flexibly model such datasets by relaxing the notion of an edge to connect an arbitrary number of vertices and providing a mathematical foundation for understanding and learning from large amounts of real-world heterogeneous data. <br> The state-of-the-art techniques for learning from graph data with pairwise relationships use graph-based deep learning models such as graph neural networks. A prominent observation that inspires this thesis is that deep neural networks are still under-explored for hypergraph data with group-wise relationships. Hypergraphs have been utilised as primary data structures in many machine learning tasks such as vertex classification, hypergraph link prediction, and knowledge base completion. However, the fundamental limitation of most existing non-neural techniques is that they cannot leverage high-dimensional features on vertices, especially those which are not present in relational data (e.g., text attributes of documents in academic networks). In this thesis, we propose novel deep learning-based methods for hypergraph data with high dimensional vertex features. <br> 1) Deep Learning for Hypergraph Vertex-level Predictions In the first part of the thesis, we focus on addressing limitations of existing methods for vertex-level tasks over hypergraphs. In particular, we propose HyperGraph Convolutional Network (HyperGCN) for semi-supervised vertex classification over hypergraphs. Unlike existing methods, HyperGCN principally bridges tools from graph neural networks and spectral hypergraph theory. <br> 2) Deep Learning for Hypergraph Link Prediction In the second part, we focus on the task of predicting groupwise relationships (i.e., link prediction over hypergraphs). We propose Neural Hyperlink Predictor (NHP), a novel neural network-based method for link prediction over hypergraphs. NHP uses a novel scoring layer that principally enables us to predict group relationships on incomplete hypergraphs where hyperedges need not represent similarity. <br> 3) Deep Learning for Multi-Relational and Recursive Hypergraphs In the third and final part, we explore more complex structures such as multi-relational hypergraphs in which each hyperedge is typed (i.e., belongs to a relation type) and recursive hypergraphs in which hyperedges can act as vertices in other hyperedges. We first propose Generalised Message Passing Neural Network (G-MPNN) for learning vertex representations on multi-relational ordered hypergraphs. G-MPNN generalises existing MPNNs on graphs, hypergraphs, multi-relational graphs, heterogeneous graphs, and multi-layer networks. We then propose MPNN-Recursive, a novel framework, to handle recursively structured data. Extensive experimentation on real-world hypergraphs shows the effectiveness of our proposed models. <br> <br> Microsoft teams Link to Online Defence: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NTk2NjE1YTYtMGNjYS00NzE0LTkxZTEtYTJhZDFkMzViZTAx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2258c2a810-9762-463d-90d0-20832b3d1592%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NTk2NjE1YTYtMGNjYS00NzE0LTkxZTEtYTJhZDFkMzViZTAx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2258c2a810-9762-463d-90d0-20832b3d1592%22%7d</a> DTSTART:20211214T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211213T120000Z UID:73bd9996f2400cdb685c802e39c1e7f2-231 DTSTAMP:19700101T120011Z DESCRIPTION:Hadwigers conjecture and total coloring. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/231/hadwigers-conjecture-and-total-coloring/ SUMMARY:Hadwigers conjecture is one of the most important and long-standing conjectures in graph theory. It was established recently that proving Hadwigers conjecture is equivalent to proving the conjecture for the special class of graphs that can be expressed as the square of some other graph. However, it is difficult to prove Hadwigers conjecture even for the squares of highly specialised graph classes. We decided to investigate the squares of subdivided graphs (a subdivided graph is a graph that can be obtained from another graph H by replacing each edge uv of H by a path uwv, where w is a new vertex). It turns out that squares of subdivided graphs are exactly the class of total graphs, well-studied in the context of the total coloring conjecture, another well-known and long-standing conjecture in graph theory. The total graph of G, denoted byT(G), is defined on the vertex set V(G) ⊔ E(G), with c1,c2 ∈ V(G) ⊔ E(G) adjacent whenever c1 and c2 are adjacent to or incident on each other in G. The total-chromatic number χ′′(G) of a graph G is defined to be equal to the chromatic number of its total graph. That is, χ′′(G) = χ(T(G)). The total coloring conjecture or TCC states that for every graph G, χ′′(G) ≤ ∆(G) + 2. Though Hadwigers conjecture was proved for the closely related class of line graphs 20 years ago itself, Hadwigers conjecture is not yet studied in the context of total graphs. In this thesis, the following results are proved : <br> (i) There exists a constant C such that, if the connectivity of G ≥ C, then Hadwigers conjecture is true for T(G). <br> (ii) Let F be a class of graphs that is closed under the operation of taking subgraphs. If a weaker version of the total coloring conjecture (weak TCC) namely, χ′′(G) ≤ ∆(G) + 3, is true for the class F, then Hadwigers conjecture is true for the class {T(G) : G ∈ F}. <br> The second statement motivates one to look for classes of graphs that satisfy weak TCC. It may be noted that a complete proof of TCC for even 4-colorable graphs (in fact even for planar graphs) has remained elusive even after decades of effort; but weak TCC can be proved easily for 4-colorable graphs. It was noticed that in spite of interest in studying χ′′(G) in terms of χ(G) right from the initial days, weak TCC is not proven to be true for k-colorable graphs, even for k= 5. In the latter half of the thesis, an important contribution to the total coloring literature is made by proving that χ′′(G) ≤ ∆(G) + 3, for every 5-colorable graph G. Being close to some of the well studied topics in total coloring, it seems that this is a long-pending addition to the literature. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWEyMmNiMWUtNzY1ZS00ZmY3LTkyYmUtNGVjOTNkNGU0NWQx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22ed2e5ccd-b870-455c-862e-26e9ab1908be%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWEyMmNiMWUtNzY1ZS00ZmY3LTkyYmUtNGVjOTNkNGU0NWQx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22ed2e5ccd-b870-455c-862e-26e9ab1908be%22%7d</a> DTSTART:20211213T120000Z END:VEVENT BEGIN:VEVENT DTEND:20211220T120000Z UID:5450377d561409722e717defe3d4b4bd-232 DTSTAMP:19700101T120011Z DESCRIPTION:MPCLeague: Robust MPC Platform for Privacy-Preserving Machine Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/232/mpcleague-robust-mpc-platform-for-privacy-preserving-machine-learning/ SUMMARY:In the modern era of computing, machine learning tools have demonstrated their potential in vital sectors, such as healthcare and finance, to derive proper inferences. The sensitive and confidential nature of the data in such sectors raises genuine concerns for data privacy. This motivated the area of Privacy-preserving Machine Learning (PPML), where privacy of data is guaranteed. Typically, machine learning techniques require significant computing power, which leads clients with limited infrastructure to rely on the method of Secure Outsourced Computation (SOC). In the SOC setting, the computation is outsourced to a set of specialized and powerful cloud servers and the service is availed on a pay-per-use basis. In this thesis, we design an efficient platform, MPCLeague, for PPML in the SOC setting using Secure Multi-party Computation (MPC) techniques. <br> MPC, the holy-grail problem of secure distributed computing, enables a set of n mutually distrusting parties to perform joint computation on their private inputs in a way that no coalition of t parties can learn more information than the output (privacy) or affect the true output of the computation (correctness). While MPC, in general, has been a subject of extensive research, the area of MPC with a small number of parties has drawn popularity of late mainly due to its application to real-time scenarios, efficiency and simplicity. This thesis focuses on designing efficient MPC frameworks for 2, 3 and 4 parties, with at most one corruption and supports ring structures. <br> <br> Our platform aims at achieving the most substantial security notion of robustness, where the honest parties are guaranteed to obtain the output irrespective of the behaviour of the corrupt parties. A robust protocol prevents the corrupt parties from repeatedly causing the computations to rerun, thereby upholding the trust in the system. While on the roadmap to attain robustness, our frameworks also demonstrate constructions with improved performance that achieve relaxed notions of security: security with abort and fairness. A fair protocol enforces the restriction that either all parties or none of them receive the output. On the other hand, honest parties may not receive the output while corrupt parties do for the case of security with abort. <br> <br> The general structure of the computation involves the execution of the protocol steps once the participating parties have supplied their inputs. Finally, the output is distributed to all the parties. However, to enhance practical efficiency, many recent works resort to the preprocessing paradigm, which splits the computation into two phases; a preprocessing phase where input-independent (but function-dependent), computationally heavy tasks can be computed, followed by a fast online phase. Since the same functions in ML are evaluated several times, this paradigm naturally fits the case of PPML, where the ML algorithm is known beforehand. <br> <br> At the heart of this thesis are four frameworks - ASTRA, SWIFT, Tetrad, ABY2.0 - catered to different settings. <br> <br> - ASTRA: We begin with the setting of 3 parties (3PC), which forms the base case for honest majority. If a majority of the participating parties are honest, then the setting is deemed an honest majority setting. In the set of 3 parties, at most one party can be corrupt, and this framework tackles semi-honest corruption, where the corrupt party follows the protocol steps but tries to glean more information from the computation. ASTRA acts as a stepping stone towards achieving a stronger security guarantee against active corruption. Our protocol requires communication of 2 ring elements per multiplication gate during the online phase, attaining a per-party cost of less than one element. This is achieved for the first time in the regime of 3PC. <br> <br> - SWIFT: Designed for 3 parties, this framework tackles one active corruption where the corrupt party can arbitrarily deviate from the computation. Building on ASTRA, SWIFT provides a multiplication that improves the communication by at least 3x over state of the art, besides improving security from abort to robustness. In the regime of malicious 3PC, SWIFT is the first robust and efficient PPML framework. It achieves a dot product protocol with communication independent of the vector size for the first time. <br> <br> - Tetrad: Designed for 4 parties in the honest majority, the fair multiplication protocol in Tetrad requires communication of only 5 ring elements instead of 6 in the state-of-the-art. The fair framework is then extended to provide robustness without inflating the costs. A notable contribution is the design of the multiplication protocol that supports on-demand applications where the function to be computed is not known in advance. <br> <br> - ABY2.0: Moving on to the stronger corruption model where a majority of the parties can be corrupt, we explore the base case of 2 parties (2PC). Since we aim to achieve robustness which is proven to be impossible in active corruption, we restrict ourselves to semi-honest corruption. The prime contribution of this framework is the scalar product for which the online communication is two ring elements irrespective of the vector dimension. This is a feature achieved for the first time in the 2PC literature. <br> <br> <br> Our frameworks provide the following contributions in addition to the ones mentioned above. First, we support multi-input multiplication for arithmetic and boolean worlds, improving the online phase in rounds and communication. Second, all our frameworks except SWIFT, incorporate truncation without incurring any overhead. Finally, we introduce efficient instantiation of garbled-world, tailor-made for the mixed-protocol framework for the first time. The mixed-protocol approach, combining arithmetic, boolean and garbled style computations, has demonstrated its potential in several practical use-cases like PPML. To facilitate the computation, we also provide the conversion mechanisms to switch between the computation styles. <br> The practicality of our framework is argued through improvements in the benchmarking of widely used ML algorithms -- Linear Regression, Logistic Regression, Neural Networks, and Support Vector Machines. We propose two variants for each of our frameworks, with one variant aiming to minimise the execution time while the other focuses on the monetary cost. <br> The concrete efficiency gains of our frameworks coupled with the stronger security guarantee of robustness make our platform an ideal choice for a real-time deployment of privacy-preserving machine learning techniques. <br> <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MDgwNTIyZmQtM2M1NC00ZGNjLTg4NjItMWNlYWM4NzliM2Fl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22a787cc01-57cc-4fc1-b7e1-4e9d51923f6d%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MDgwNTIyZmQtM2M1NC00ZGNjLTg4NjItMWNlYWM4NzliM2Fl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22a787cc01-57cc-4fc1-b7e1-4e9d51923f6d%22%7d</a> DTSTART:20211220T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220103T120000Z UID:7c773c24574476d56eda7d969b17c1e1-233 DTSTAMP:19700101T120017Z DESCRIPTION:Operating System and Hypervisor Support for Mitigating the Address Translation Wall URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/233/operating-system-and-hypervisor-support-for-mitigating-the-address-translation-wall/ SUMMARY:Virtual memory has proven to be an extremely powerful abstraction for its programmability benefits. Unfortunately, virtual memory is becoming a performance bottleneck due to the address translation wall. Modern applications with large memory footprints necessitate frequent page table walks to perform the virtual to physical address translation. Consequently, the hardware spends 30-50% of the total CPU cycles in servicing TLB misses alone. Virtualization and non-uniform memory access (NUMA) architectures further exacerbate this overhead. For example, virtualized systems involve two-dimensional page table walks that require up to 24 memory accesses for each TLB miss, with current 4-level page tables. The address translation performance drops further on NUMA systems, depending on the distance between the CPU and page tables. These overheads will increase in the future, where deeper page tables and multi-tiered memory systems will enable even larger applications. Virtual memory, therefore, is showing its age in the era of data-centric computing. <br> This thesis investigates the role of an operating system (OS) and hypervisor in improving the address translation performance. First, we focus on huge pages that can significantly reduce the frequency and cost of TLB misses. Huge pages are widely available in modern systems e.g., x86 architecture supports 2MB and 1GB huge pages, in addition to regular 4KB pages. While huge pages are great in theory, real-world OSs have often delivered disappointing performance while using them. This is because memory management of huge pages is fraught with multiple challenges. We propose several enhancements in OS-level policies and mechanisms to make huge pages beneficial, even under multi-dimensional constraints such as latency, capacity, and fairness. <br> Second, we investigate the effect of NUMA on address translation performance. NUMA architectures mandate careful data placement to hide the effect of variable memory access latency from applications. Several decades of research on NUMA systems have optimized access to user-level application data. However, prior research has ignored the access performance of kernel data, including page tables, due to their small memory footprint. We argue that it is time to revisit page table management for NUMA-like systems. <br> <br> The core contributions of this thesis include four systems: Illuminator, HawkEye, Trident, and vMitosis, as summarized below: <br> Illuminator: We first expose some subtle implications of external memory fragmentation on huge pages. We show that despite proactive measures employed in the memory management subsystem of Linux, unmovable kernel objects (e.g., inodes, page tables, etc.) can deny huge pages to user applications. In a long-running system, unmovable objects fragment physical memory, often permanently, and cause high de-fragmentation overheads. Over time, their effects manifest in performance regressions, OS jitter, and latency spikes. Illuminator effectively clusters kernel objects in a subset of physical memory regions and makes huge page allocations feasible even under heavily fragmented scenarios.. <br> HawkEye: In this work, we deal with OS-based huge page management policies that need to balance complex trade-offs between TLB coverage, memory bloat, latency, and the number of page faults. In addition, we consider performance and fairness issues that appear under fragmentation when memory contiguity is limited. In HawkEye, we propose asynchronous page pre-zeroing to simultaneously optimize for low latency and few page faults. We propose automatic bloat recovery to effectively deal with the trade-offs between TLB coverage and memory bloat at runtime. HawkEye addresses the performance and fairness challenges by allocating huge pages based on their estimated profitability. <br> <br> Trident: Illuminator and HawkEye try to extract maximum benefits from 2MB huge pages. However, recent findings have shown that even after employing 2MB pages, more than 20% of the total CPU cycles are wasted in handling TLB misses for data center applications. We address this problem using 1GB huge pages that provide up to 1TB per-core TLB coverage on modern systems. Leveraging insights from our earlier work, we propose a multi-level huge page framework called Trident that judiciously allocates 1GB, 2MB, and 4KB pages as deemed suitable at runtime. <br> vMitosis: In this work, we focus on the effect of NUMA on address translation in virtualized servers. We show that page table walks often involve remote memory accesses on NUMA systems that can slow down large memory applications by more than 3x. Interestingly, the slow down observed due to remote page table accesses can even outweigh that of accessing remote data, even though page tables consume less than 1% memory of overall application footprint. vMitosis mitigates the effect of NUMA on page table walks by enabling each core to handle TLB misses from its local socket. We achieve this by judiciously migrating and replicating page tables across NUMA sockets. <br> <br> Overall, with this thesis, we show that adequate OS and hypervisor support can help virtual memory thrive even in the era of data-centric computing. We have implemented our proposed systems in the Linux OS kernel and KVM hypervisor. Our optimizations are transparent to the users, and using them does not require any hardware or application modifications. <br> Teams link: <br><a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjA2OGFmMmYtNjBjMi00NDJlLTk2NTItZjI5YTBlOTY3Yjgw%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2282f39501-c5b2-4bfb-87c3-f17ca74c00b6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjA2OGFmMmYtNjBjMi00NDJlLTk2NTItZjI5YTBlOTY3Yjgw%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2282f39501-c5b2-4bfb-87c3-f17ca74c00b6%22%7d</a> DTSTART:20220103T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220107T120000Z UID:f43c6769d541a3ca9ad1a2815de788da-234 DTSTAMP:19700101T120017Z DESCRIPTION:Playing Unique Games on Certifiable Small-set Expanders URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/234/playing-unique-games-on-certifiable-small-set-expanders/ SUMMARY:The Unique Games Conjecture (UGC) is a central open question in computational complexity and algorithms. In short, the UGC stipulates that distinguishing between almost satisfiable and highly unsatisfiable instances of a certain 2-variable constraint satisfaction problem (CSP) called Unique Games is NP-hard. We build algorithms for UG on a large class of restricted instances called certifiable small-set expanders. In doing so we give new tools to analyze Sum-of-Squares SDPs and connect the UGC to another important complexity theoretic conjecture, the Small-Set-Expansion Hypothesis. The talk will start from a basic introduction to the UG problem and will not assume prior knowledge about UG or sum-of-squares algorithms. <br> <br> This is based on joint work - https://arxiv.org/abs/2006.09969 <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220107T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220111T120000Z UID:acc273db1acdf3ac7c7827bb0ec0dab7-235 DTSTAMP:19700101T120010Z DESCRIPTION:Analysis and Methods for Knowledge Graph Embeddings URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/235/analysis-and-methods-for-knowledge-graph-embeddings/ SUMMARY:Knowledge Graphs (KGs) are multi-relational graphs where nodes represent entities, and typed edges represent relationships among entities. These graphs store real-world facts such as (Lionel Messi, plays-for-team, Barcelona) as edges, called triples. KGs such as NELL, YAGO, Freebase, and WikiData have been very popular and support many AI applications such as Web Search, Query Recommendation, and Question Answering. Although popular, these KGs suffer from incompleteness. Learning Knowledge Graph Embeddings (KGE) is a popular approach for predicting missing edges (i.e., link prediction) and representing entities and relations in downstream tasks. While numerous KGE methods have been proposed in the past decade, our understanding and analysis of such embeddings have been limited. Further, such methods only work well with ontological KGs. In this thesis, we address these gaps. <br> Firstly, we study various KGE methods and present an extensive analysis of these methods, resulting in interesting insights. Next, we address an under-explored problem of link prediction in Open Knowledge Graphs (OpenKGs) and present a novel approach that improves the type compatibility of predicted edges. Lastly, we present an adaptive interaction framework for learning KG embeddings that generalizes many existing methods. <br> Analysis of KGE Embeddings In the first part, we present a macro and a micro analysis of embeddings learned by various KGE methods. <br> Despite the popularity and effectiveness of KG embeddings, their geometric understanding (i.e., arrangement of entity and relation vectors in vector space) is unexplored. We initiate a study to analyze the geometry of KG embeddings and correlate it with task performance and other hyper-parameters. Firstly, we present a set of metrics (e.g., Conicity, ATM) to analyze the geometry of a group of vectors. Using these metrics, we find sharp differences between the geometry of embeddings learned by different classes of KGE methods. The vectors learned by a multiplicative model lie in a narrow cone, unlike additive models where the vectors are spread out in the space. This behavior of multiplicative models is amplified by increasing the number of negative samples used for training. Further, a very high Conicity value is negatively correlated with the performance on the link prediction task. <br> We also study the problem of understanding KG embeddings semantics and propose an approach to learn more coherent dimensions. A dimension is coherent if the top entities have similar types (e.g., person). In this work, we formalize the notion of coherence using entity co-occurrence statistics and propose a regularizer term that maximizes coherence while learning KG embeddings. The proposed approach significantly improves coherence while having a comparable performance with baseline in the link prediction and triple classification tasks. Further, based on the human evaluation, we demonstrate that the proposed approach learns more coherent dimensions than the baseline. <br> <br> A method for OpenKG Embedding In the second part, we address the problem of learning KG embeddings for Open Knowledge Graphs (OpenKGs), focusing on improving link prediction. An OpenKG refers to a set of (head noun phrase, relation phrase, tail noun phrase) triples such as (tesla, return to, new york) extracted from a text corpus using OpenIE tools. While OpenKGs are easy to bootstrap for a domain, they are very sparse. Therefore, link prediction becomes an important step while using these graphs in downstream tasks. <br> Learning OpenKG embeddings is one approach for link prediction that has received some attention lately. However, on careful examination, we find that current algorithms often predict noun phrases (NPs) with incompatible types for given noun and relation phrases. We address this problem and propose OKGIT that improves OpenKG link prediction using novel type compatibility score and type regularization. With extensive experiments on multiple datasets, we show that the proposed method achieves state-of-the-art performance while producing type compatible NPs in the link prediction task. <br> An Adaptive Framework for KG Embeddings In the third part, we address the problem of improving the KGE models. <br> Firstly, we show that the performance of existing approaches varies across different datasets, and a simple neural network-based method can consistently achieve better performance on these datasets. Upon analysis, we find that KGE models depend on fixed sets of interactions among the dimensions of entity and relation vectors. <br> Therefore, we investigate ways to learn such interactions automatically during training. We propose an adaptive interaction framework for learning KG embeddings, which can learn appropriate interactions while training. We show that some of the existing models could be seen as special cases of the proposed framework. Based on this framework, we also present two new models, which outperform the baseline models on the link prediction task. Further analysis demonstrates that the proposed approach can adapt to different datasets by learning appropriate interactions. <br> <br> Microsoft teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NmVhNDk3ZDctOWVlNC00MWE0LWEzMTItNWIyNDE3YTlhNDFh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2258c2a810-9762-463d-90d0-20832b3d1592%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NmVhNDk3ZDctOWVlNC00MWE0LWEzMTItNWIyNDE3YTlhNDFh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2258c2a810-9762-463d-90d0-20832b3d1592%22%7d</a> DTSTART:20220111T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220121T120000Z UID:9c9f0b879f8b33d9dbfab346366b2f16-237 DTSTAMP:19700101T120017Z DESCRIPTION:Planar Partition Oracles and When they strike gold: An exp(1/epsilon^2) Tester for All Planar Properties URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/237/planar-partition-oracles-and-when-they-strike-gold-an-exp1-epsilon2-tester-for-all-planar-properties/ SUMMARY:Take a planar graph with maximum degree d. These graphs admit a hyperfinite decompositions, where, for a sufficiently small ϵ > 0, one removes ϵdn edges to get connected components of size independent of n. An important tool for sublinear algorithms and property testing for such classes is the partition oracle. A partition oracle is a local procedure that gives consistent access to a hyperfinite decomposition, without any preprocessing. Given a query vertex v, the partition oracle outputs the component containing v in time independent of n. All the answers are consistent with a single hyperfinite decomposition. In this talk, I will describe a partition oracle that runs in time poly(d/ϵ). I will also describe how this machinery strikes a fortune and helps in obtaining a constant time tester for all planar properties. This is easily obtained by a better error analysis on the seminal result of Newman and Sohler [SICOMP 2013]. Based on Joint works with Sabyasachi Basu, C. Seshadhri and Andrew Stolman. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220121T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220119T120000Z UID:bc13ca13e3d7b2e96ce4c3046951f3c5-238 DTSTAMP:19700101T120014Z DESCRIPTION:Achieving practical secure non-volatile memory system with in-Memory Integrity Verification (iMIV) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/238/achieving-practical-secure-non-volatile-memory-system-with-in-memory-integrity-verification-imiv/ SUMMARY:Recent commercialization of Non-Volatile Memory (NVM) technology in the form of Intel Optane enables programmers to write recoverable programs. However, the data on NVM is susceptible to a plethora of data remanence attacks, which makes confidentiality and integrity protection of data essential for a secure NVM system. However, that requires computing and maintaining a large amount of security metadata (encryption counters, message authentication code (MAC), and integrity tree nodes (BMT)). Furthermore, crash consistency guarantees require the system to persist the security metadata and data atomically back to NVM, incurring high overheads. So there is a trade-off between providing complete security guarantees, the performance and recovery time of an NVM system. <br> To ensure the confidentiality and integrity of data, a substantial quantity of security metadata is required. Of these, persisting Bonsai Merkel Tree (BMT) nodes, which are essential for fine-grain integrity verification, add substantial cost owing to the massive amount of data that must be moved off-chip to the bandwidth-constrained NVM. Thus, prior works often make a trade-off between performance and fine-grain verifiability or forego it entirely in favor of performance. <br> The goal of this thesis is to maintain strong security and verifiability guarantees while limiting the cost of BMT updates and my thesis accomplishes this by leveraging the in-memory integrity verification. We make the fine-grain integrity verifiability realizable with a radically different approach of using in-memory computing for integrity verification. Our proposal, iMIV draws inspiration from the fact that today's commercial Optane NVM performs encryption onboard the DIMM. We argue that memory-intensive integrity verification operation should be performed near the (non-volatile) memory to avoid off-chip data movement. iMIV targets to minimize the off-chip memory transfer and mitigate the effect of bandwidth wall and also scales to larger NVM capacity in future systems with per-DIMM BMT. <br> The experiments and analysis are carried out on a trace-driven cycle-accurate simulator VANS, which mimics the internal micro-architecture of Intel Optane memory DIMMs. The experimental results show that in comparison to the Baseline scheme with write-through caches and strict persistency model, which also provides complete security guarantees, iMIV reduces system runtime by 1.8 times for persistent-memory aware workloads and 3.4 times for persistent-memory agnostic workloads. iMIV's recovery time on system crashes is microseconds-scale without compromising on detecting tampering and fast pin-point of the unverifiable region. iMIV limits the performance overheads associated with fine-grain integrity verifiability to less than 55 percent compared to a system that offers no security. DTSTART:20220119T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220127T120000Z UID:d76017e6fb5ac3a664e4a163e9c27bf0-239 DTSTAMP:19700101T120015Z DESCRIPTION:Neural Models for Personalized Recommendation Systems with External Information URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/239/neural-models-for-personalized-recommendation-systems-with-external-information/ SUMMARY:Personalized recommendation systems use the data generated by user-item interactions (for example, in the form of ratings) to predict different users interests in available items and recommend a set of items or products to the users. The sparsity of data, cold start, and scalability are some of the important challenges faced by the developers of recommendation systems. These problems are alleviated by using external information, which can be in the form of a social network or a heterogeneous information network, or cross-domain knowledge. This thesis develops novel neural network models for designing personalized recommendation systems using the available external information. <br> The first part of the thesis studies the top-N item recommendation setting where the external information is available in the form of a social network or heterogeneous information network. Unlike a simple recommendation setting, capturing complex relationships amongst entities (users, items, and connected objects) becomes essential when a social and heterogeneous information network is available. In a social network, all socially connected users do not have equal influence on each other. Further, estimating the quantum of influence among entities in a user-item interaction network is important when only implicit ratings are available. We address these challenges by proposing a novel neural network model, SoRecGAT, which employs a multi-head and multi-layer graph attention mechanism. The attention mechanism helps the model learn the influence of entities on each other more accurately. Further, we exploit heterogeneous information networks (HIN) to gather multiple views for the items. A novel neural network model -- GAMMA (Graph and Multi-view Memory Attention mechanism) is proposed to extract relevant information from HINs. The proposed model is an end-to-end model which eliminates the need for learning a similarity matrix offline using some manually selected meta-paths before optimizing the desired objective function. <br> In the second part of the thesis, we focus on top-N bundle recommendation and list continuation setting. Bundle recommendation is the task of recommending a group of products instead of individual products to users. We study two interesting challenges -- (1) how to personalize and recommend existing bundles to users and (2) how to generate personalized novel bundles targeting specific users. We propose GRAM-SMOT -- a graph attention-based framework that considers higher-order relationships among the users, items, and bundles and the relative influence of items present in the bundles. For efficiently learning the embeddings of the entities, we define a loss function based on the metric-learning approach. A strategy that leverages submodular optimization ideas is used to generate novel bundles. <br> We also study the problem of top-N personalized list continuation where the task is to curate the next items to user-generated lists (ordered sequence of items) in a personalized way by using the sequential information of the items in the list. The main challenge in this task is understanding the ternary relationships among the users, items, and lists. We propose HyperTeNet -- a self-attention hypergraph and Transformer-based neural network architecture for the personalized list continuation task. Here, graph convolutions are used to learn the multi-hop relationship among entities of the same type. A self-attention-based hypergraph neural network is proposed to learn the ternary relationships among the interacting entities via hyperlink prediction in a 3-uniform hypergraph. Further, the entity embeddings are shared with a Transformer-based architecture and are learned through an alternating optimization procedure. <br> The final part of the thesis focuses on the personalized rating prediction setting where external information is available in the form of cross-domain knowledge. We propose an end-to-end neural network model, NeuCDCF, that provides a way to alleviate data sparsity problems by exploiting the information from related domains. NeuCDCF is based on a wide and deep framework and learns the representations jointly using matrix factorization and deep neural networks. We study the challenges involved in handling diversity between domains and learning complex non-linear relationships among entities within and across domains. <br> We conduct extensive experiments in each of these settings using several real-world datasets and demonstrate the efficacy of the proposed models. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NjRmMDg2OTktMjM5ZC00NjMyLWFiZmMtNmQ1NGY2NTkzNGU2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NjRmMDg2OTktMjM5ZC00NjMyLWFiZmMtNmQ1NGY2NTkzNGU2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20220127T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220128T120000Z UID:1211f97ae7d04e882e6bfee2ddf540a8-240 DTSTAMP:19700101T120011Z DESCRIPTION:Matrix Discrepancy from Quantum Communication URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/240/matrix-discrepancy-from-quantum-communication/ SUMMARY:In this talk, we will discuss a novel connection between discrepancy minimization and (quantum) communication complexity. As an application, we resolve a substantial special case of the Matrix Spencer conjecture. In particular, we show that for every collection of $n$ $n times n$ symmetric matrices $A_1 dots A_n$ with spectral norm bounded by 1 and Frobenius norm bounded by$n^{1/4}$, there is a signing $x$ such that $|| sum x_i A_i|| leq sqrt{n}$ We give a polynomial-time algorithm based on partial coloring and semidefinite programming to find such a sign. The proof of our main result combines a simple compression scheme for transcripts of repeated (quantum) communication protocols with quantum state purification, the Holevo bound from quantum information, and tools from sketching and dimensionality reduction. Our approach also offers a promising avenue to resolve the Matrix Spencer conjecture completely -- we show it is implied by a natural conjecture in quantum communication complexity. <br> The talk is based on joint work with Sam Hopkins (MIT) and Prasad Raghavendra (UC Berkeley). <br> Microsoft Teams Link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220128T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220131T120000Z UID:6940bd8c2fd14351f5ffea9e97ff5332-241 DTSTAMP:19700101T120015Z DESCRIPTION:Explainable and Efficient Neural models for Natural Language to Bash Command Translation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/241/explainable-and-efficient-neural-models-for-natural-language-to-bash-command-translation/ SUMMARY:One of the key goals of Natural Language Processing is to make computers understand natural language. Semantic Parsing has been one of the driving tasks for Natural Language Understanding. It is formally defined as the task of generating meaning representation from natural language input. In this work, we focus on using the Bash command as the meaning representation. Bash is a Unix command language used for interacting with the Operating System. Recent works on natural language to Bash command translation have made significant advances on this problem. The best performing solutions employ a neural network architecture called the Transformer. In this work, we explore the aspects of explainability and efficiency for this task and use the Transformer as one of the baselines for comparing the proposed approaches. <br> In the first part, we utilize documentation data from Linux manual pages and the Abstract Syntax Tree for Bash to generate explanations for the translated Bash command. We propose a novel architecture that incorporates tree structure information in the Transformer and provides explanations for its predictions via alignment matrices between user invocation and manual page text. We find that the proposed method performs on par with the Transformer performance. Our method performs better than fine-tuned T5, a Transformer-based neural model pre-trained on a large amount of text data in a self-supervised manner. <br> In the second part, we use the problems inherent synchronous structure and propose the Segmented Invocation Transformer (SIT) that utilizes the information from the constituency parse tree of the natural language invocation. Our method is motivated by the alignment between segments in the natural language text and Bash command components. By utilizing this structure, the proposed method outperforms the state-of-the-art approach while achieving a 1.8x improvement in the inference time (as measured on a CPU) and a 5x reduction in model parameters. We also conduct an attribution analysis using Integrated Gradients to empirically confirm the identified structure and the ability of SIT to capture it. <br> Microsoft team link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWJhNTU3ZTYtYzQ2YS00OGI4LWE3ZjktZTYxMmNjNDNiYTdi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWJhNTU3ZTYtYzQ2YS00OGI4LWE3ZjktZTYxMmNjNDNiYTdi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20220131T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220204T120000Z UID:81c5f81ec8cffa825e36d4884dc020f8-242 DTSTAMP:19700101T120011Z DESCRIPTION:Cascaded Norms in Clustering URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/242/cascaded-norms-in-clustering/ SUMMARY:Clustering is one of the most fundamental tasks in various areas such as machine learning and optimization. In theoretical computer science, we are interested in the complexity of finding a good clustering, given a data set with some distance function, and a target number of centers to choose from among the input points. Our goal is to find a set of centers (of the required cardinality) which minimizes some cost function which aggregates the distances of all input points from their respective nearest centers. This includes well-studied notions such as k-Medians Clustering and k-Means Clustering. <br> More recently, there has been a focus on fairness in clustering, in which we want to take into consideration not only the global cost but also to counteract structural bias against marginalized groups. To this end, one first aggregates the costs incurred within the given groups of interest, before aggregating the costs incurred by these groups. <br> We focus on a very general notion of fairness - the input consists of data points, a target number of centers, and a collection of groups represented by different weight functions. The objective we wish to minimize is the L_q norm of the group costs, where each group cost is computed as the (weighted) L_p norm of distances of points in the group to their respective nearest centers. We study this problem from the point of view of approximation algorithms, giving algorithms for all values of p and q that smoothly interpolate between optimal and near-optimal approximations for fundamental parameter settings of (p,q), such as (infinity, q), (p, infinity), and (p,p). <br> <br> Based on joint work with Yury Makarychev and Ali Vakilian. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220204T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220204T120000Z UID:593094f86d5da4dfe132b8719e85f584-243 DTSTAMP:19700101T120016Z DESCRIPTION:Fast Algorithms for Max Cut on Geometric Intersection Graphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/243/fast-algorithms-for-max-cut-on-geometric-intersection-graphs/ SUMMARY:In the max cut problem, given a graph, the goal is to partition the vertex set into two disjoint sets such that the number of edges having their endpoints in different sets is maximized. Max cut is an NP-hard problem. The seminal work by Goemans and Williamson gave an approximation algorithm for the max cut problem having an approximation ratio of 0.878. <br> In this work, we design fast algorithms for max cut on geometric intersection graphs. In a geometric intersection graph, given a collection of n geometric objects as the input, each object corresponds to a vertex and there is an edge between two vertices if and only if the corresponding objects intersect. <br> Since we are dealing with the geometric intersection graphs, which have more structure than general graphs, the following questions are of interest: 1. Are there special cases of geometric intersection graphs for which max cut can be solved exactly in polynomial time? 2. It can be shown that the random cut gives a 0.5 approximation for the max cut. Is it possible to design linear or near-linear time algorithms (in terms of n) and beat the 0.5 approximation barrier? The edge-set of the graph is not explicitly given as input; therefore, designing linear time algorithms is of interest. 3. Can an approximation factor better than 0.878 be obtained for the geometric intersection graphs? <br> We obtain the following results: An exact and fast algorithm for laminar geometric intersection graphs. Our algorithm uses a greedy strategy. A fast algorithm is obtained by combining the properties of laminar objects with range searching data structures. <br> An O(n log n) time algorithm with an approximation factor of 2/3 for unit interval intersection graphs. We decompose the unit intervals into several cliques, and based on the number of edges between adjacent cliques, we choose an appropriate partitioning strategy. <br> An O(n log n) time algorithm with an approximation factor of 7/13 for unit square intersection graphs. We use the "largest clique" in the graph to beat the 0.5 approximation barrier. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZDk5YmM2YjctZGU0Ni00NTg3LTkwNDAtZjVlYmRjOGE1Mjdk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d71644d4-9e9d-48e9-ab41-3a147899b402%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZDk5YmM2YjctZGU0Ni00NTg3LTkwNDAtZjVlYmRjOGE1Mjdk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d71644d4-9e9d-48e9-ab41-3a147899b402%22%7d</a> DTSTART:20220204T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220209T120000Z UID:3300a59022d2f7a714fd35341e35af73-244 DTSTAMP:19700101T120008Z DESCRIPTION:On Software Regression Test Parallelization URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/244/on-software-regression-test-parallelization/ SUMMARY:Software contributes significantly to our daily activities, ranging from small-scale basic utilities to large-scale safety-critical systems. With ever changing requirements comes the need for a mandatory software update operation to avoid disruption of service. However, the problem arises when unwanted bugs occur due to the update and pose a threat to the expected functionality of the software. To mitigate this issue, software testing calls for a systematic approach called regression testing. Software is said to have progressed if the new version is bug-free, otherwise it is said to have regressed. The regression testing component of the software development life cycle comprises several sub-components that can benefit from parallelization. In this talk, I mention my contributions towards the literature of software regression testing powered by multi-core parallelization. We overview challenges that prohibit efficient parallelization, and in depth discuss a lightweight approach that provides soundy automated parallel test-execution. <br> Teams Meeting link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTVlMWM4ZjEtMmZmYy00OTE0LWJkYzMtMDUwYTEyNDUyZGZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OTVlMWM4ZjEtMmZmYy00OTE0LWJkYzMtMDUwYTEyNDUyZGZl%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20220209T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220215T120000Z UID:7b81424d4826173c049aab9c4fb0e4cf-245 DTSTAMP:19700101T120015Z DESCRIPTION:Algorithms for Fair Clustering URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/245/algorithms-for-fair-clustering/ SUMMARY:Many decisions today are taken by various machine learning algorithms, hence it is crucial to accommodate fairness in such algorithms to remove/reduce any kind of bias in the decision. We incorporate fairness in the problem of clustering. Clustering is a classical machine learning problem in which the task is to partition the data points into various groups such that the data points belonging to one group are more similar to each other than the data points belonging to some other group in the partition. In our model, each data point belongs to one or more number of categories. We define fairness in terms of two constraints, restricted dominance and minority protection. While ensuring fairness in the clustering, we consider each data point in only one of the categories from the set of categories it belongs to. Our model ensures that no category is either in minority or in dominance in any of the clusters. Representation of a category in a cluster is considered not in absolute terms but in proportion to its presence in the whole dataset. We give bi-criteria approximation for fair clustering whose objective is to minimise $L_p$-norm. Here, the $L_p$-norm is defined as begin{align*} L_p(V,phi) = left(sum_{v in V} d(v,phi(v))^pright)^{1/p}, end{align*} where $V$ is the dataset, $C$ is the set of centers chosen for clustering, $Phi:V rightarrow C$ is the assignment which minimises the cost of clustering while satisfying the fairness constraints and $p$ can take any positive integral value. Our solution violates the fairness constraints by an additive violation of at most $2$. We implement this algorithm and do experiments to compare it with the state-of-the-art. For any $epsilon >0$, we give a $(1 + epsilon)$-approximate algorithm for fair clustering for points lying in Euclidean space whose objective is to minimise $L_1$-norm (or $L_2$-norm). This algorithm also violates the fairness constraints by an additive violation of at most $2$. For points lying in $mathbb{R}^d$, the run time of this algorithm for $L_2$-norm is $Oleft(nd cdot 2^{tilde{O}(k/epsilon)}right) + poly(n) cdot 2^{tilde{O}(k/epsilon)}$, where $n$ represents the size of the dataset. For $L_1$-norm, the run time of this algorithm is $Oleft(nd cdot 2^{tilde{O}(k/epsilon^{O(1)})}right) + poly(n) cdot 2^{tilde{O}(k/epsilon^{O(1)})}$. Given a $gamma$-perturbation resilient instance of clustering in the metric space $(V,d)$, we also give a bi-criteria approximation for the fair clustering of the same instance while changing its metric to $d $. Here, $d $ is any metric which is a $gamma$-perturbation of $(V,d)$. This solution also violates the fairness constraints by an additive violation of at most $2$. <br> <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDIzMjlkYWQtNDY1Yi00ZmMwLTk3OTktMTI5YTY5NWRkNzg0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220c3c2a63-37e3-4ad6-b0bd-ddfa9589e2d5%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDIzMjlkYWQtNDY1Yi00ZmMwLTk3OTktMTI5YTY5NWRkNzg0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%220c3c2a63-37e3-4ad6-b0bd-ddfa9589e2d5%22%7d</a> DTSTART:20220215T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220211T120000Z UID:ba2bd34db3a7dd8433f002579de1b0bc-246 DTSTAMP:19700101T120011Z DESCRIPTION:Improving Reliability and Performance of Datacenter Systems via Coherence URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/246/improving-reliability-and-performance-of-datacenter-systems-via-coherence/ SUMMARY:Reliability and performance are key metrics for modern datacenter machines. Co-designing for these introduces delicate trade-off decisions for system architects. In this talk I present 2 works, where we aim to improve both reliability and performance of modern shared memory hardware in the datacenter by designing tailored coherence protocols. <br> In the first work, we aim to combat increased memory system failure rates. We propose Dvé, a hardware-driven replication mechanism where data blocks are replicated in 2 different sockets across a cache-coherent NUMA system. Such an organization has the advantage of offering two independent points of access to data which enables: (a) strong error correction that can recover from a range of faults affecting any of the components in the memory, upto and including the memory controller, and (b) higher performance by providing another nearer point of memory access. Dvé realizes both of these benefits via Coherent Replication, a technique that builds on top of existing cache coherence protocols. Coherent Replication keeps the replicas in sync for reliability and provides coherent access to the replicas during fault-free operation for performance. Dvé introduces a unique design point that offers higher reliability and performance flexibly on-demand. <br> In the second work, we propose to improve reliability and performance of function-as-a-service (FaaS) deployments. The FaaS model allows applications to be decomposed into a workflow of stand-alone functions which are instantiated and executed on-demand in the cloud. The stateless nature of this model forces functions to store/retrieve data from a remote object store, thereby adding latency. Our work Bolt, uses all-hardware memory disaggregation to build an object store for FaaS applications. Bolt builds on top of the latest cache-coherent attachment technologies for off-chip memory peripherals like GenZ, CXL or NVLink2 to enable an all-hardware solution. It adds an object granularity  caching mechanism to cache objects in hardware caches at compute nodes, hence improving performance of FaaS functions. Bolt then adds an inter-node cache coherence mechanism that ensures the data in the compute node caches is consistent. <br> Bolt’s coherence ensures reliable operation in such a loosely coupled system by providing an asynchronous, non-blocking protocol which ensures forward progress during partial system failures. <br> Teams Meeting Link: <br> <a href="https://tinyurl.com/AdarshPatilsTalk">https://tinyurl.com/AdarshPatilsTalk</a> DTSTART:20220211T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220218T120000Z UID:87a99217e21b8e10f55123febb55cec0-247 DTSTAMP:19700101T120011Z DESCRIPTION:Parameterized Approaches to Kemeny Rank Aggregation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/247/parameterized-approaches-to-kemeny-rank-aggregation/ SUMMARY:The Kemeny aggregation method defines a consensus ranking based on a collection of rankings. It relies on the Kendall-tau distance between two rankings, which is defined as the number of pairs that are ranked differently. An optimal ranking in this context is one that minimizes the total distance from each of the rankings in the given collection. In this talk, we survey some early lines of work that approached this problem from a parameterized perspective, and its connections with the Feedback Arc Set problem. We then discuss some recent developments, showcasing the use of structural parameterizations for finding not just one solution, but a diverse set of solutions. <br> <br> Most of this talk will be based on the paper Diversity in Kemeny Rank Aggregation: A Parameterized Approach, by Emmanuel Arrighi, Henning Fernau, Daniel Lokshtanov, Mateus de Oliveira Oliveira, Petra Wolf (IJCAI 2021). <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220218T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220217T120000Z UID:f84b4b1674953ef158f1954604631081-248 DTSTAMP:19700101T120015Z DESCRIPTION:Multi-Armed Bandits – On Range Searching and On Slowly-varying Non-stationarity. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/248/multi-armed-bandits-ae-on-range-searching-and-on-slowly-varying-non-stationarity/ SUMMARY:Multi-Armed Bandits (MAB) is a popular framework for modelling sequential decision-making problems under uncertainty. <br> This thesis is a compilation of two independent works on MABs. <br> 1. Optimal Algorithms for Range Searching over Multi-Armed Bandits We consider a multi-armed bandit (MAB) version of the range-searching problem. In its basic form, range searching considers as input a set of points (on the real line) and a collection of (real) intervals. Here, with each specified point, we have an associated weight, and the problem objective is to find a maximum-weight point within every given interval. <br> The current work addresses range searching with stochastic weights: each point corresponds to an arm (that admits sample access) and the points weight is the (unknown) mean of the underlying distribution. In this MAB setup, we develop sample-efficient algorithms that find, with high probability, near-maximum-weight points (arms) within the given intervals, i.e., we obtain PAC (probably approximately correct) guarantees. We also provide an algorithm for a generalization wherein the weight of each point is a multi-dimensional vector. The sample complexities of our algorithms depend, in particular, on the size of the optimal hitting set of the given intervals. <br> Finally, we establish lower bounds proving that the obtained sample complexities are essentially tight. Our results highlight the significance of geometric constructs -- specifically, hitting sets -- in our MAB setting. <br> <br> 2. On Slowly-varying Non-stationary Bandits We consider minimisation of dynamic regret in non-stationary bandits with a slowly varying property. Namely, we assume that arms rewards are stochastic and independent over time, but that the absolute difference between the expected rewards of any arm at any two consecutive time-steps is at most a drift limit δ>0. For this setting that has not received enough attention in the past, we give a new algorithm which extends naturally the well-known Successive Elimination algorithm to the non-stationary bandit setting. We establish the first instance-dependent regret upper bound for slowly varying non-stationary bandits. The analysis in turn relies on a novel characterization of the instance as a detectable gap profile that depends on the expected arm reward differences. We also provide the first minimax regret lower bound for this problem, enabling us to show that our algorithm is essentially minimax optimal. Also, this lower bound we obtain matches that of the more general total variation-budgeted bandits problem, establishing that the seemingly easier former problem is at least as hard as the more general latter problem in the minimax sense. We complement our theoretical results with experimental illustrations. <br> <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YTZkNTU0ZWYtMzAxZS00NjZjLWFkZjktMzI0M2M0OTg5OTA3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c747ccaa-ceaa-4197-b4cb-ce2f1d4694da%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YTZkNTU0ZWYtMzAxZS00NjZjLWFkZjktMzI0M2M0OTg5OTA3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c747ccaa-ceaa-4197-b4cb-ce2f1d4694da%22%7d</a> DTSTART:20220217T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220228T120000Z UID:d1893ecd6ac0687234ebe98eb24d354e-250 DTSTAMP:19700101T120016Z DESCRIPTION:Online Learning with Markovian Data via Reverse Experience Replay URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/250/online-learning-with-markovian-data-via-reverse-experience-replay/ SUMMARY:Learning with Markovian data is a challenging problem with several applications in critical domains like reinforcement learning, control theory, time series analysis etc. Techniques like SGD are the workhorse of large-scale learning, with rigorous analysis for streaming i.i.d. data in several regimes. But their applicability to non i.i.d. Markovian data is unclear due to dependency between points.<br> In this talk, we will present results showing that SGD in general can be significantly sub-optimal for Markovian data. In contrast, through three critical problems: a) linear regression, b) dynamical system identification aka vector auto-regressive model estimation, c) policy learning with Linear MDPs, we demonstrate that SGD enhanced with experience replay--a popular heuristic used in RL literature--leads to nearly optimal solutions. To the best of our knowledge, we provide the first rigorous analysis of the practically popular experience replay technique. Similarly, our result provides the first provably efficient Q-learning style method for finding optimal policy for linear MDPs.<br> Based on joint works with Naman Agarwal, Syomantak Chaudhuri, Suhas Kowshik, Dheeraj Nagaraj, Praneeth Netrapalli, Carrie Wu. <br> <br> After the talk we will then have a 30 mins Q&A session with students to make them aware of various opportunities at Google Research India including predoc roles and student researchership.<br> -------------------------------------------------------------------------<br> Google research India has several opportunities for graduating/final year students to participate in research projects. Here is a partial list:<br> 1. Predoc researcher: Candidates who hold Bachelors/Masters degrees can spend up to two years working with research teams at Google research India on cutting-edge research projects. Most of our past pre-doc researchers have then gone on to pursue PhDs at top schools such as MIT, Berkeley, CMU, University of Washington etc.<br> 2. Student researcher/ student internships: This is for students who are in their final/pre-final year of their Bachelors/Masters or any year of their PhD program to spend part of their time working on research projects with teams at Google research India. DTSTART:20220228T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220303T120000Z UID:bf487493199c3ebce6b922abb39d82d7-251 DTSTAMP:19700101T120018Z DESCRIPTION:Electrical Flows over Spanning Trees URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/251/electrical-flows-over-spanning-trees/ SUMMARY:The network reconfiguration problem seeks to find a rooted tree T such that the energy of the (unique) feasible electrical flow over T is minimized. The tree requirement on the support of the flow is motivated by operational constraints in electricity distribution networks. The bulk of existing results on convex optimization over vertices of polytopes and on the structure of electrical flows do not easily give guarantees for this problem, while many heuristic methods have been developed in the power systems community as early as 1989. Our main contribution is to give the first provable approximation guarantees for the network reconfiguration problem. We provide novel lower bounds and corresponding approximation factors for various settings ranging from min(O(m-n), O(n)) for general graphs, to O(sqrt{n}) over grids with uniform resistances on edges, and O(1) for grids with uniform edge resistances and demands. To obtain the result for general graphs, we propose a new method for (approximate) spectral graph sparsification, which may be of independent interest. Using insights from our theoretical results, we propose a general heuristic for the network reconfiguration problem that is orders of magnitude faster than existing methods in the literature, while obtaining comparable performance. <br> <br> Microsoft Teams Link: <br> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220303T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220309T120000Z UID:9c23afecd9d9141e72a8dc656dd45234-252 DTSTAMP:19700101T120016Z DESCRIPTION:Top-k Spatial Aware Ads URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/252/top-k-spatial-aware-ads/ SUMMARY:Consider an app on a smartphone which displays local business ads. When a user opens the app, then k local business ads need to be displayed (where k would typically be 3 or 5) such that the profit made by the app is maximized. The pricing model needs to consider that (a) each business is willing to bid a different price, and (b) farther the distance of the user on whose smartphone the ad is displayed, the lesser is the price paid by to the app. <br> Motivated by such applications, in this work, we design fast algorithms to retrieve top-k ads using spatial and non-spatial information. We refer to them as Top-k Spatial Aware Ads Queries (SAA). In Top-k SAA, we return top-k objects that have the best score, and the scoring function is based on the distance between the object and query point (spatial attribute) and non-spatial attributes. We propose data structures that efficiently preprocess the input data and aid in fast query processing. A simple O(n log n) algorithm sorts the ads based on the scoring function value and returns the top-k ads. In addition, R-Trees can be used with some modification as data structure to answer these queries. All these methods have high worst-case bounds. In this thesis our goal is to obtain better worst-case bounds in terms of space and query time. We obtain the following results.<br> <br> Our first algorithm uses O(nlog n) space and answers the top-k SAA query in O(k log2n) time. <br> The fast query time is obtained by leveraging the properties of additively weighted Voronoi diagram, along with other supporting data structures. Our second algorithm improves upon the first algorithm by improving the query time to O(k log n), while using the same space. This is achieved via an interesting combination of randomization with a top-2 structure. <br> The proposed algorithms have good worst-case theoretical bounds. We demonstrate, through experiments, that our algorithms perform well in terms of query time and space. <br> Microsoft teams link:<br> <a href="https://tinyurl.com/25vcrmpy">https://tinyurl.com/25vcrmpy</a> DTSTART:20220309T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220304T120000Z UID:d536135b4e9b09a4b5a4b79ab7a7d46f-253 DTSTAMP:19700101T120016Z DESCRIPTION:Reinforcement Learning Via Sequence Modeling URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/253/reinforcement-learning-via-sequence-modeling/ SUMMARY:I will introduce a framework that abstracts Reinforcement Learning (RL) as a sequence modeling problem. This allows us to draw upon the simplicity and scalability of the Transformer architecture, and associated advances in language modeling such as GPT-x and BERT. I will present Decision Transformer (DT), an architecture that casts the problem of RL as conditional sequence modeling. Unlike prior approaches to RL that fit value functions or compute policy gradients, DT simply outputs the optimal actions by leveraging a causally masked Transformer. By conditioning an autoregressive model on the desired return (reward), past states, and actions, DT can generate future actions that achieve the desired return. I will also present our recent work proposing entropy regularizers to extend DT to online learning with hindsight learning and entropy-based regularization. Despite its simplicity, Decision Transformer matches or exceeds the performance of state-of-the-art model-free offline and online RL baselines on benchmark environments. DTSTART:20220304T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220311T120000Z UID:de9f416ac27dc93ee691dbc3f3d470e7-255 DTSTAMP:19700101T120010Z DESCRIPTION:Online Learning with Hints URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/255/online-learning-with-hints/ SUMMARY:We consider the online learning problem, where at every step the algorithm makes a decision x_t and incurs a loss given by a loss function ell_t, which is then revealed to the algorithm. The goal is to achieve low regret, which is defined as the difference between the cumulative loss of the algorithm, and the total loss of the best fixed decision in hindsight. Classical results in the area show that a sublinear regret can be achieved for a wide class of loss functions. <br> I will talk about a line of recent work in which we assume that the learner has access to a hint about the loss function at every step. For instance, in the setting of online linear optimization where ell_t(x) is simply the inner product for some cost vector c_t, a hint can correspond to a vector that is "mildly correlated" with c_t. In such settings, we show that one can significantly improve upon known regret bounds. We show that our algorithms can deal with hints occasionally being "bad" (uncorrelated or misleading), and also work in settings where we can only ask for hints in a small number of time steps. <br> Most of the talk is joint work with Ashok Cutkosky (Boston University), Ravi Kumar and Manish Purohit (Google Research). <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220311T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220311T120000Z UID:38167c4a89f6af048827d7cedefec1d4-256 DTSTAMP:19700101T120010Z DESCRIPTION:Learning Invariants for Verification of Programs and Control Systems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/256/learning-invariants-for-verification-of-programs-and-control-systems/ SUMMARY:Learning Invariants for Verification of Programs and Control Systems <br> Abstract: <br> Deductive verification techniques in the style of Floyd and Hoare have the potential to give us concise, compositional, and scalable proofs of the correctness of various kinds of software systems like programs and control systems. However, the major hurdle in adopting these techniques is that the verification engineer often needs to come up with adequate program invariants (like loop invariants), which may require a lot of expertise and manual effort. We try to address this problem by attempting to automate the process of finding adequate invariants, using machine learning and other techniques. <br> In the first part of the presentation, we introduce a data-driven learning-based technique to automate the computation of adequate invariants for the deductive verification of various classes of programs, including recursive sequential programs, and concurrent programs. We consider standard pre-post specifications for these programs. Deductive verification of programs can be viewed as finding a solution to a system of Constrained Horn Clauses (CHCs) induced by the program and its specification. Earlier works on data-driven learning of invariants like ICE learning (Garg et al, 2014) can only handle linear CHCs. However, many deductive verification techniques like Requires-Ensures (for recursive sequential programs) and Rely-Guarantee and Owicki-Gries (for concurrent programs) induce non-linear CHCs. We propose a technique called Horn-ICE, which extends the ICE learning technique to solve non-linear CHCs, thereby allowing us to automate the deductive verification of a variety of new classes of programs. <br> Non-linear CHCs give rise to Horn implications counterexamples (Horn clauses) in general, so our learning setup considers a semi-labelled dataset along with its metadata in the form of Horn clauses. We have implemented a decision tree based classifier for this semi-labelled dataset. This classifier learns candidate invariants for a deductive verification task. We demonstrate the performance of this automated program verification technique using standard benchmarks like SV-COMP. On the sequential benchmark suite, our tool is on par with the state-of-the-art tool Z3/PDR. On the recursive benchmark suite, we outperform the state-of-the-art tool Ultimate Automizer (Heizmann et al, 2013). On the concurrent benchmark suite, we are able to verify some popular programs within a few seconds. <br> In the second part of the presentation, we consider sampled-data control systems that control a continuous plant using a discrete controller, and the band convergence property for these systems. Our first contribution in this part is a deductive verification technique for sampled-data control systems. This technique symbolically executes the system from the initial state onwards until an adequate invariant is reached. For this verification technique, we developed a concolic execution based white-box approach for learning adequate invariants. We demonstrate the performance of our automated sampled-data control system verification technique using standard Simulink models. Our toolchain is able to verify band convergence properties for most of these models within a few seconds. <br> Details for joining online on Google Meet: <br> Link: <a href="https://meet.google.com/gvp-wzwd-tcp">https://meet.google.com/gvp-wzwd-tcp</a> DTSTART:20220311T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220317T120000Z UID:9bc40b8992b5e88b8e04db6f624f478e-257 DTSTAMP:19700101T120011Z DESCRIPTION:Recovery Algorithms for planted structures in Semi-random models URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/257/recovery-algorithms-for-planted-structures-in-semi-random-models/ SUMMARY:For many NP-hard problems, the analysis of best-known approximation algorithms yield “poor” worst-case guarantees. However, using various heuristics, the problems can be solved (to some extent) in real-life instances. This success can be attributed to the atypicality of worst-case instances in real life, and therefore motivates studying the problem in “easier” instances. Analyzing the problem in Planted solution models and Semi-random models is one such systematic approach along these lines. <br> <br> In this thesis, we study planted solution models and semi-random models for various graph problems. Given a graph G with n vertices, we consider the task of finding the largest induced subgraph of G with a particular structure. We start by studying the problem where the particular structure is a planar graph. Next, we look at the Odd Cycle Transversal problem or equivalently the problem of finding the largest induced bipartite subgraph. Finally, we study the problem of finding the largest independent set in r-uniform hypergraphs. All these problems are NP-hard and have abysmal worst-case approximation guarantees. <br> An instance of a planted solution model is constructed by starting with a set of vertices V, and choosing a set S ⊆ V of k vertices, and adding a particular structure to it. Edges between pairs of vertices in S x (VS) and (VS) x (VS) are added independently with probability p. The algorithmic task then is to recover this planted structure. As a special case for all these problems, when the planted structure is an empty graph, the problem reduces to recovering a planted independent set and we dont expect recovery algorithms for k =o(n^{1/2}). <br> For the problem of finding the largest induced bipartite subgraph, we give an exact recovery algorithm that works for k = Ω_p((n.log n)^{1/2}). For the problem of finding the maximum independent set in r-uniform hypergraphs, we give an algorithm that works for Ω_p(n^{(r-1)/(r-0.5)}). Our results also hold for a natural semi-random model of instances inspired by Feige and Kilian [FK01] model. Our algorithms are based on analyzing continuous relaxations of these problems. We employ techniques from Spectral Graph Theory, Convex Optimization (Linear Programs (LPs) and Semi-Definite Programs (SDPs) relaxations), and Lasserre/Sum-of-Squares hierarchy strengthening of convex relaxations. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MmQ3ODc2Y2YtMjZhZS00ODQzLTk4ZDgtMjhjNTY1YTcxYjE3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a2358f-870a-4a9b-aaed-f1fbcc8c2f75%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MmQ3ODc2Y2YtMjZhZS00ODQzLTk4ZDgtMjhjNTY1YTcxYjE3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2220a2358f-870a-4a9b-aaed-f1fbcc8c2f75%22%7d</a> DTSTART:20220317T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220317T120000Z UID:fa31a5c3e06cbd1302ba5516aa9da194-258 DTSTAMP:19700101T120021Z DESCRIPTION:Partitioning over Submodular Structures – Part I URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/258/partitioning-over-submodular-structures-ae-part-i/ SUMMARY:In submodular k-partitioning problems, the input is a submodular set function (given via an evaluation oracle) along with a fixed positive integer k (e.g., k = 2, 3, 4, …) and the goal is to partition the ground set into k non-empty parts in order to minimize certain natural objectives of interest. Submodular k-partitioning generalizes partitioning problems over several interesting structures including graphs and hypergraphs. The case of 2-partitioning corresponds to the classic and well-studied submodular minimization problem which is polynomial-time solvable. In this talk, I will survey some recent progress towards polynomial-time solvability of submodular k-partitioning for fixed constants k>=3. As a main technical result, I will present a random contraction algorithm for min-cut in hypergraphs – this is a generalization of Kargers algorithm for min-cut in graphs. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220317T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220315T120000Z UID:7f9e391ea245afbbfd1362370b1205f6-259 DTSTAMP:19700101T120019Z DESCRIPTION:Structured Encryption and Dynamic Leakage Suppression URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/259/structured-encryption-and-dynamic-leakage-suppression/ SUMMARY:Structured encryption (STE) schemes encrypt data structures in such a way that they can be privately queried. Special cases of STE include searchable symmetric encryption (SSE) and graph encryption. Like all sub-linear encrypted search solutions, STE leaks information about queries against persistent adversaries. To address this, a line of work on leakage suppression was recently initiated that focuses on techniques to mitigate the leakage of STE schemes. A notable example is the query equality suppression framework (Kamara et al. CRYPTO18) which transforms dynamic STE schemes that leak the query equality into new schemes that do not. Unfortunately, this framework can only produce static schemes and it was left as an open problem to design a solution that could yield dynamic constructions. In this work, we propose a dynamic query equality suppression framework that transforms volume-hiding semi-dynamic or mutable STE schemes that leak the query equality into new fully-dynamic constructions that do not. We then use our framework to design three new fully-dynamic STE schemes that are “almost” and fully zero-leakage which, under natural assumptions on the data and query distributions, are asymptotically more efficient than using black-box ORAM simulation. These are the first constructions of their kind. This is joint work with Seny Kamara (Brown University) and Tarik Moataz (Aroki Systems). DTSTART:20220315T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220324T120000Z UID:b3963a6bd87ce45d775a48866e76871e-260 DTSTAMP:19700101T120021Z DESCRIPTION:Partitioning over Submodular Structures – Part II URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/260/partitioning-over-submodular-structures-ae-part-ii/ SUMMARY:In submodular k-partitioning problems, the input is a submodular set function (given via an evaluation oracle) along with a fixed positive integer k (e.g., k = 2, 3, 4, …) and the goal is to partition the ground set into k non-empty parts in order to minimize certain natural objectives of interest. Submodular k-partitioning generalizes partitioning problems over several interesting structures including graphs and hypergraphs. The case of 2-partitioning corresponds to the classic and well-studied submodular minimization problem which is polynomial-time solvable. In this talk, I will present a polynomial time algorithm for minmax symmetric submodular k-partitioning for every fixed k. Microsoft Teams Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220324T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220401T120000Z UID:5b5f08cab5a70420f368d04271272077-261 DTSTAMP:19700101T120010Z DESCRIPTION:An Evaluation of Basic Protection Mechanisms in Financial Apps on Mobile Devices URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/261/an-evaluation-of-basic-protection-mechanisms-in-financial-apps-on-mobile-devices/ SUMMARY:Mobile devices have become an integral part of the payment ecosystem. Payments are facilitated by financial applications (like Mobile Banking, UPI Apps, etc.), which have in turn soared in popularity. With the increasing dependence on the financial app ecosystem and the sensitive nature of the data handled by financial apps (including the bank/card details of the payees and the payers), we set out to study fundamental question: do the app developers of financial apps put various self-defense checks to make their apps more secure? If yes, how trivial is it for the attackers to bypass such checks? <br> This thesis concerns the robustness of security checks in financial mobile applications. The best practices recommended by the Open Web Application Security Project (OWASP) for developing such apps, demand that developers include several checks in these apps, such as detection of running on a rooted device, certificate checks, and so on. Ideally, these checks must be introduced in a sophisticated way and must not be locatable through trivial static analysis, so that attackers cannot bypass them trivially. In this work, we conduct a large-scale study focused on financial apps on the Android platform and determine the robustness of these checks. <br> Our study shows that a significant fraction of the financial apps dont have the various self-defense checks recommended by the OWASP. Then we showed that among the apps with at least one security check, > 50% of such apps at least one check could be trivially bypassed. Some of such financial apps have installation counts exceeding 100 million from Google Play. This entire process of detecting the self-defense check and bypassing it is automated. We believe that the results of our study can guide developers of these apps in inserting security checks in a more robust fashion. DTSTART:20220401T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220401T120000Z UID:58f1df3544640fb880bb93a8ad40880e-262 DTSTAMP:19700101T120014Z DESCRIPTION:The effect of network delays on Distributed Ledgers based on Direct Acyclic Graphs: A mathematical model URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/262/the-effect-of-network-delays-on-distributed-ledgers-based-on-direct-acyclic-graphs-a-mathematical-model/ SUMMARY:In this talk, we introduce a new mathematical model to analyse the performance of fully distributed ledgers based on Directed Acyclic Graphs (DAG), under the presence of heterogeneous delay. In particular, we focus on the IOTA foundations tangle, while our results can be extended to other DAG-based distributed ledgers. State-of-the-art mathematical models trying to capture the impact of delays on the performance of such distributed ledgers rely on some particular approximations. In contrast, through our model, we are able to analytically derive similar performance guarantees, in a more realistic set-up. We contrast our results with results obtained in a real testbed, showing good accordance between them. <br> Microsoft Teams <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzU0NTlmYzEtOTU2OC00YTdjLWJkZmUtY2YzMDRlNzk4ZDJm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22adc1e56f-56ee-4d24-873f-341c97ae782a%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzU0NTlmYzEtOTU2OC00YTdjLWJkZmUtY2YzMDRlNzk4ZDJm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22adc1e56f-56ee-4d24-873f-341c97ae782a%22%7d</a> DTSTART:20220401T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220404T120000Z UID:cc01baaef2aa2516c9a914c8541a87d4-263 DTSTAMP:19700101T120015Z DESCRIPTION:2-Level Page Tables (2-LPT): A building block for efficient address translation in virtualized environment URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/263/2-level-page-tables-2-lpt-a-building-block-for-efficient-address-translation-in-virtualized-environment/ SUMMARY:Efficient address translation mechanisms are gaining more and more attention as the virtual address range of the processors keeps expanding and the demand for machine virtualization increases with cloud and data center-based services.  Traditional radix-tree based address translations can incur significant overheads in big data applications, particularly under virtualization, due to multi-level tree walks and nested translation. The overheads stem primarily from unnecessary generality --- ability to support several hundreds of thousands of virtual memory regions in the virtual address space --- supported by current processors.   <br> We observe that in the common case, however, a process's virtual address space contains only a few contiguously allocated sections, which can be efficiently translated using a shallow tree with two levels. We propose such a compact structure, called 2-Level Page Table(2-LPT),  which is a key building block for address translation in virtualized environment. A key advantage of 2-LPT is that it maintains two levels of page tables irrespective of the size of the virtual address space. Translating a virtual address (VA) using 2-LPT is fast. A walk on a 2-LPT requires up to two memory accesses. In practice, however,  the root level table is well cached in the PWC, thus, single memory access is sufficient. Under native execution, 2-LPT reduces the time spent in page walks by up to 20.9% (9.38% on average) and improves performance by up to 10.1% (1.66% on average) over the conventional four-level radix tree page tables, on a set of memory-intensive applications. <br> 2-LPT is more beneficial under virtualization. The proposed 2-LPT design reduces the cost of nested page walk from 24 to 8 memory accesses.  To achieve further reduction, we propose two optimizations: (i)  Enhanced Partial Shadow Paging (ePSP) which employs a limited form of shadow paging for the root-level of 2-LPT, and (ii) Host PTE Mirroring (HPM) which allows accessing the host page table entry without performing host page table walk. These allow us to largely avoid slow VM exits while effectively reducing the number of memory access on a nested address translation to just one, on average. 2-LPT speeds up applications by 5.6%-50.9% (24.4%, on average) over the baseline with conventional nested page walks. Importantly, it reduces page walk cycles and execution time of the best performing state-of-the-art proposal by 17.1%-57.1% and by 3.9%-43.9%, respectively. <br> <br> Online Meeting Link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWU2Y2IxZDAtMGVkNC00NmUwLWJkMjMtZWVhNTk0ZmYyODEx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWU2Y2IxZDAtMGVkNC00NmUwLWJkMjMtZWVhNTk0ZmYyODEx%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%224bcd3d56-e405-4b06-99fb-27742262f261%22%7d</a> DTSTART:20220404T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220401T120000Z UID:e10fe120aeb93579c13d5bb5b71eb930-264 DTSTAMP:19700101T120014Z DESCRIPTION:Model-based Safe Deep Reinforcement Learning and Empirical Analysis of Safety via Attribution URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/264/model-based-safe-deep-reinforcement-learning-and-empirical-analysis-of-safety-via-attribution/ SUMMARY:During initial iterations of training in most Reinforcement Learning (RL) algorithms, agents perform a significant number of random exploratory steps, which in the real-world limit the practicality of these algorithms as this can lead to potentially dangerous behavior. Hence safe exploration is a critical issue in applying RL algorithms in the real world. This problem is well studied in the literature under the Constrained Markov Decision Process (CMDP) Framework, where in addition to single-stage rewards, state transitions receive single-stage costs as well. The prescribed cost functions are responsible for mapping undesirable behavior at any given time-step to a scalar value. Then we aim to find a feasible policy that maximizes reward returns and keeps cost returns below a prescribed threshold during training as well as deployment. <br> We propose a novel On-policy Model-based Safe Deep RL algorithm in which we learn the transition dynamics of the environment in an online manner as well as find a feasible optimal policy using Lagrangian Relaxation-based Proximal Policy Optimization. This combination of transition dynamics learning, and a safety-promoting RL algorithm leads to ~3-4 times less environment interactions and less cumulative hazard violations compared to the model-free approach. We use an ensemble of neural networks with different initializations to tackle epistemic and aleatoric uncertainty issues faced during environment model learning. We present our results on a challenging Safe Reinforcement Learning benchmark - the Open AI Safety Gym. <br> In addition to this, we perform an attribution analysis of actions taken by the Deep Neural Network-based policy at each time step. This analysis helps us to : <br> 1. Identify the feature in state representation which is significantly responsible for the current action. <br> 2.Empirically provide the evidence of the safety-aware agents ability to deal with hazards in the environment provided that hazard information is present in the state representation. In order to perform the above analysis, we assume state representation has meaningful information about hazards and goals. Then we calculate an attribution vector of the same dimension as state using a well-known attribution technique known as Integrated Gradients. The resultant attribution vector provides the importance of each state feature for the current action. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzU5MGY4YTctNThmYi00NjlkLTllZmItNDc5ZjExMzY2ZTU4%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2271844033-661c-432d-9a6f-418de5b8c819%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzU5MGY4YTctNThmYi00NjlkLTllZmItNDc5ZjExMzY2ZTU4%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2271844033-661c-432d-9a6f-418de5b8c819%22%7d</a> DTSTART:20220401T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220408T120000Z UID:7a3e42ff72d19e00f757ba5c6811687f-265 DTSTAMP:19700101T120015Z DESCRIPTION:Reinforcement Learning Algorithms for Off-Policy, Multi-Agent Learning and Applications to Smart Grids URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/265/reinforcement-learning-algorithms-for-off-policy-multi-agent-learning-and-applications-to-smart-grids/ SUMMARY:Reinforcement Learning (RL) algorithms are a popular class of algorithms for training an agent to learn desired behavior through interaction with an environment whose dynamics is unknown to the agent. RL algorithms combined with neural network architectures have enjoyed much success in various disciplines like games, medicine, energy management, economics, and supply chain management. In our thesis, we study interesting extensions of standard single-agent RL settings, like off-policy and multi-agent settings. We discuss the motivations and importance of these settings and propose convergent algorithms to solve these problems. Finally, we consider one of the important applications of RL, namely smart grids. The goal of the smart grid is to develop a power grid model that intelligently manages its energy resources. In our thesis, we propose RL models for efficient smart grid design. <br> Learning the value function of a given policy (target policy) from the data samples obtained from a different policy (behavior policy) is an important problem in Reinforcement Learning (RL). This problem is studied under the setting of off-policy prediction. Temporal Difference (TD) learning algorithms are a popular class of algorithms for solving prediction problems. TD algorithms with linear function approximation are convergent when the data samples are generated from the target policy (known as on-policy prediction) itself. However, it has been well established in the literature that off-policy TD algorithms under linear function approximation may diverge. In the first part of the thesis, we propose a convergent online off-policy TD algorithm under linear function approximation. The main idea is to penalize updates of the algorithm to ensure convergence of the iterates. We provide a convergence analysis of our algorithm. Through numerical evaluations, we further demonstrate the effectiveness of our proposed scheme. <br> Subsequently, we consider the ``off-policy control setup in RL, where an agents objective is to compute an optimal policy based on the data obtained from a behavior policy. As the optimal policy can be very different from the behavior policy, learning optimal behavior is very hard in the ``off-policy setting compared to the ``on-policy setting wherein the data is collected from the new policy updates. In this work, we propose the first deep off-policy natural actor-critic algorithm that utilizes state-action distribution correction for handling the off-policy behavior and the natural policy gradient for sample efficiency. Unlike the existing natural gradient-based actor-critic algorithms that use only fixed features for policy and value function approximation, the proposed natural actor-critic algorithm can utilize a deep neural networks power to approximate both policy and value function. We illustrate the benefit of the proposed off-policy natural gradient algorithm by comparing it with the Euclidean gradient actor-critic algorithm on benchmark RL tasks. <br> In the third part of the thesis, we consider the problem of two-player zero-sum games. In this setting, there are two agents, both of whom aim to optimize their payoffs. Both the agents observe the same state of the game, and the agents objective is to compute a strategy profile that maximizes their payoffs. However, the payoff of the second agent is the negative of the payoff obtained by the first agent. Therefore, the objective of the second agent is to minimize the total payoff obtained by the first agent. This problem is formulated as a min-max Markov game in the literature. In this work, we compute the solution of the two-player zero-sum game utilizing the technique of successive relaxation. Successive relaxation has been successfully applied in the literature to compute a faster value iteration algorithm in the context of Markov Decision Processes. We extend the concept of successive relaxation to the two-player zero-sum games. We then derive a generalized minimax Q-learning algorithm that computes the optimal policy when the model information is unknown. Finally, we prove the convergence of the proposed generalized minimax Q-learning algorithm utilizing stochastic approximation techniques. Through experiments, we demonstrate the advantages of our proposed algorithm. <br> Next, we consider a cooperative stochastic games framework where multiple agents work towards learning optimal joint actions in an unknown environment to achieve a common goal. In many real-world applications, however, constraints are often imposed on the actions that the agents can jointly take. In such scenarios, the agents aim to learn joint actions to achieve a common goal (minimizing a specified cost function) while meeting the given constraints (specified via certain penalty functions). Our work considers the relaxation of the constrained optimization problem by constructing the Lagrangian of the cost and penalty functions. We propose a nested actor-critic solution approach to solve this relaxed problem. In this approach, an actor-critic scheme is employed to improve the policy for a given Lagrange parameter update on a faster timescale as in the classical actor-critic architecture. Using this faster timescale policy update, a meta actor-critic scheme is employed to improve the Lagrange parameters on the slower timescale. Utilizing the proposed nested actor-critic scheme, we develop three Nested Actor-Critic (N-AC) algorithms. <br> In recent times, actor-critic algorithms with attention mechanisms have been successfully applied to obtain optimal actions for RL agents in multi-agent environments. In the fifth part of our thesis, we extend this algorithm to the constrained multi-agent RL setting considered above. The idea here is that optimizing the common goal and satisfying the constraints may require different modes of attention. Thus, by incorporating different attention modes, the agents can select useful information required for optimizing the objective and satisfying the constraints separately, thereby yielding better actions. Through experiments on benchmark multi-agent environments, we discuss the advantages of our proposed attention-based actor-critic algorithm. <br> In the last part of our thesis, we study the applications of RL algorithms to Smart Grids. We consider two important problems - on the supply-side and demand-side, respectively, and study both in a unified framework. On the supply side, we study the problem of energy trading among microgrids to maximize profit obtained from selling power while at the same time satisfying the customer demand. On the demand side, we consider optimally scheduling the time-adjustable demand - i.e., of loads with flexible time windows in which they can be scheduled. While previous works have treated these two problems in isolation, we combine these problems and provide a unified Markov decision process (MDP) framework for these problems. <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzEzMmI3MTUtOGUxMS00ZWZiLTgwZjAtMDcwMTI4ZDYzMzRh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d931ce1d-4f82-48c0-8053-96272991d288%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzEzMmI3MTUtOGUxMS00ZWZiLTgwZjAtMDcwMTI4ZDYzMzRh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d931ce1d-4f82-48c0-8053-96272991d288%22%7d</a> DTSTART:20220408T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220407T120000Z UID:b843f4b85b43ec3f3b18e1e7fa5e46f9-266 DTSTAMP:19700101T120016Z DESCRIPTION:Alice in the PA-Land URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/266/alice-in-the-pa-land/ SUMMARY:The speaker will take us on a wondrous journey through different variants of the Privacy Amplification problem (with different levels of Eavesdroppers power) and associated with the variants of extractors. Well start with a toy version of the problem then discuss a classical DW09 result finally culminating in a grand finale: variant with corrupted randomness sources and tampered memory from AORSS20 and COA21(order randomized). The main result includes the latest construction of a two-source non-malleable extractor, which surpasses all known constructions of non-malleable seedless/seeded extractors. We will only show the general overview of the construction/compiler. <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d ">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220407T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220412T120000Z UID:68b7feec09f3918c5a82f3307028ac0c-267 DTSTAMP:19700101T120010Z DESCRIPTION:Novel First-order Algorithms for Non-smooth Optimization Problems in Machine Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/267/novel-first-order-algorithms-for-non-smooth-optimization-problems-in-machine-learning/ SUMMARY:This thesis is devoted to designing efficient optimization algorithms for machine learning (ML) problems where the underlying objective function to be optimized is convex but not necessarily differentiable. Such non-smooth objective functions are ubiquitous in ML mainly due to the use of one or more of the following: (a) non-differentiable loss function (hinge loss in binary classification), (b) sparsity promoting regularization term (L1 norm in regression), (c) constraint sets (elliptope in graph embedding) to induce specific structure on the parameters to be learned. Motivated by such a wide range of learning problems that can be cast as optimization of a non-smooth convex objective, we focus on developing first-order algorithms with non-asymptotic convergence rate guarantees to solve such problems in a large-scale setting. Based on shortcomings of the existing research in this domain, we address the following specific issues in this thesis. <br> First, we consider the problem of minimizing a convex function over a feasible set given by the intersection of finitely many simple sets, each of which is equipped with a projection oracle. Examples of constraint sets that possess such structure include the set of doubly stochastic matrices, elliptope, the intersection of PSD cone with an L1-norm ball, etc. The main difficulty lies in computing the projection of a point onto the feasible set. Existing approaches yield an infeasible point with no guarantees on its in-feasibility. Exploiting the intersecting sets linear regularity property, we present an exact penalty approach that leads to first-order algorithms with explicit guarantees on the approximate solutions distance from the feasible set. Further, we show improved iteration-complexity when the objective possesses structural smoothness / strong convexity. This is achieved through a saddle-point reformulation where the proximal operators required by the primal-dual algorithms can be computed in closed form. We illustrate the benefits of our approach on a graph transduction problem and graph matching. <br> Next, we consider the classic composite problem of minimizing the sum of two convex functions: a smooth one (possessing Lipschitz continuous gradient) and a simple non-smooth one with easy to compute proximal operator. The well-known FISTA algorithm (also Nesterovs accelerated gradient method) achieves the optimal O(1/T^2) non-ergodic convergence rate for this problem. One of the drawbacks of these fast gradient methods is that they require computing gradients of the smooth function at points different from those on which the convergence rate guarantee applies. Inspired by Polyaks Heavy Ball method and the use of past gradients as momentum in training deep nets, we propose an accelerated gradient algorithm to rectify this drawback keeping the convergence rate intact. We achieve this through a judicious choice of momentum in both primal and dual space. To the best of our knowledge, this is the first accelerated gradient algorithm that achieves O(1/T^2) convergence rate guarantee on the iterates at which gradients are evaluated. Next, we propose modifications of Nesterovs accelerated gradient method by calling the first-order oracle at the same points on which the convergence rate guarantees apply. Algorithms thus obtained additionally achieve a linear convergence rate when the objective function is strongly convex. This fills a significant research gap as Polyaks Heavy Ball method guarantees accelerated convergence rate through gradient momentum only for a restrictive class of twice differentiable and strongly convex objective functions. <br> Third, we focus on the problem of learning a positive semidefinite (PSD) kernel matrix from m similarity matrices under a general convex loss. The existing algorithms do not apply if one employs arbitrary loss functions and often can not handle m>1 case. Based on the black-box approach of Mirror Descent (MD), we present several provably convergent iterative algorithms that exploit the availability of off-the-shelf support vector machine (SVM) solvers. One of the significant contributions involves an extension of the well-known MD algorithm for simplex to handle the Cartesian product of PSD matrices leading to a novel algorithm called Entropic Multiple Kernel Learning. We also show simulation results on protein structure classification involving several similarity matrices to demonstrate the proposed algorithms efficacy. <br> Finally, we consider the class of problems possessing convex-concave saddle point structure with bilinear interaction. This model encompasses most convex optimization problems arising in ML and includes minimizing the sum of many simple non-smooth convex functions as a special case; thereby, it subsumes learning problems involving complex regularization terms such as total-variation based image denoising, overlapping group lasso, isotonic regression, etc. We first propose a primal-dual algorithm for this general class of problems that can achieve the O(1/T) convergence rate guarantee on the non-ergodic primal-dual iterate pair. Further, assuming strong convexity in the primal domain, we show an improved non-ergodic convergence rate of O(1/T^2). In contrast, the existing primal-dual algorithms achieve such convergence rates only in the ergodic or semi-ergodic setting. Further, we propose another primal-dual algorithm utilizing an additional prox-operator computation (in the primal space) per iteration. This variant additionally enjoys a emph{non-ergodic} accelerated linear convergence rate when the objective function is strongly convex-strongly concave. The proposed algorithms were designed by cleverly incorporating the key ingredients from Nesterovs accelerated gradient method in the standard primal-dual algorithmic framework of Chambolle-Pock. <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmY5ZDc3YTctNGJiNC00MDkyLThjNTktM2RjNDcxNjQxZTVh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22620fb6db-36c5-4f95-ba45-6ed8e824aa28%22%7d"https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmY5ZDc3YTctNGJiNC00MDkyLThjNTktM2RjNDcxNjQxZTVh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22620fb6db-36c5-4f95-ba45-6ed8e824aa28%22%7d>https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmY5ZDc3YTctNGJiNC00MDkyLThjNTktM2RjNDcxNjQxZTVh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22620fb6db-36c5-4f95-ba45-6ed8e824aa28%22%7d</a> DTSTART:20220412T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220426T120000Z UID:156ac7faf90d86635a67256837ca3f20-269 DTSTAMP:19700101T120014Z DESCRIPTION:Performance Characterization of .NET Benchmarks URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/269/performance-characterization-of-net-benchmarks/ SUMMARY:Managed language frameworks are pervasive, especially in modern datacenters. .NET is one such framework that is used widely in Microsoft Azure but has received little attention from computer architects. In this work we organize a set of representative .NET benchmarks and characterize them for performance bottlenecks on modern hardware. Our study reveals that .NET applications have different characteristics than traditional programs due to the managed runtime. This affects the tradeoffs when designing hardware for such applications. We use Principal Component Analysis (PCA) and hierarchical clustering to create representative subsets of open-source .NET and ASP.NET benchmarks. We perform microarchitecture and application-level characterization of these subsets and show that they are significantly different from SPEC CPU17 benchmarks in branch and memory behavior. We also analyze the effect of managed runtime events such as JIT (Just-in-Time) compilation and GC (Garbage Collection). Among other findings, GC improves cache performance and JITing could benefit from aggressive prefetching. As computing increasingly moves to the cloud and managed languages become more popular, it is important to include .NET-like benchmarks in architecture studies. DTSTART:20220426T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220427T120000Z UID:6e82f5b799b0863917e7a9493594dc54-270 DTSTAMP:19700101T120014Z DESCRIPTION:Future of Compute URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/270/future-of-compute/ SUMMARY:Designing a high performance architecture to run continually evolving machine learning models is a complex problem. At the same time, rapid adoption of AI is increasing the need for scalable systems to run such models. This technical talk will apply first-principles thinking to system architecture that can be the foundation for AI/ML compute. We will discuss the structure of a machine learning model and how next generation AI compute uses Scalar, Vector, Matrix and Tensor architecture. We will also expand on the components that will become the backbone of AI architectures, such as new data formats and RISCV - an open source instruction set architecture. Finally we will look at the implications of building chips and systems in this new era of artificial intelligence and how it will drive Foundry, Silicon, SOC, IPs, Data Center, Cloud and Software 2.0, the next iteration in software development. The talk will be followed by a Networking and Interaction Session with the Students. DTSTART:20220427T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220429T120000Z UID:fc8f56238b78b06d8da781d15d5bb8bf-271 DTSTAMP:19700101T120016Z DESCRIPTION:Near Optimal Split-state Non-malleable Codes URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/271/near-optimal-split-state-non-malleable-codes/ SUMMARY:At ITCS 2010, Dziembowski, Pietrzak, and Wichs introduced Non-malleable Codes (NMCs) which protect against tampering of a codeword of a given message into the codeword of a related message. A well-studied model of tampering is the 2-split-state model where the codeword consists of two independently tamperable states. As with standard error-correcting codes, it is of great importance to build codes with high rates. <br> Following a long line of work, Aggarwal and Obremski (FOCS 2020) showed the first constant rate non-malleable code in the 2−split state model; however, this constant was a minuscule 10^{-6}! In our work[1], we build a Non-malleable Code with rate 1/3 (nearly matches the rate 1/2 lower bound for this model). This work will be the focus of my talk! <br> [1] Rate One-Third Non-malleable Codes, STOC 2022. Divesh Aggarwal, Sruthi Sekar, Bhavana Kanukurthi, Maciej Obremski, Sai Lakshmi Bhavana Obbattu <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220429T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220428T120000Z UID:fe61626207dc116923372cbf6234655a-272 DTSTAMP:19700101T120011Z DESCRIPTION:Neural Approaches for Natural Language Query Answering over Source Code URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/272/neural-approaches-for-natural-language-query-answering-over-source-code/ SUMMARY:During software development, developers need to ensure that the developed code is bug-free and the best coding practices are followed during the code development process. To guarantee this, the developers require answers to queries about specific aspects of the code relevant to the development. Powerful code-query languages such as CodeQL have been developed for this purpose. Use of such code-query languages, however, requires expertise in writing formal queries. For each separate query, one needs to write several lines in a code-query language. <br> To remedy these problems, we propose to represent each query by a natural language phrase and answer such queries using neural networks. We aim to perform model training such that a single model can answer multiple queries as opposed to writing separate formal queries for each task. Such a model can answer these queries against unseen code. With this motivation, we introduce the novel NlCodeQA dataset. It includes 171,346 labeled examples where each input consists of a natural language query and a code snippet. The labels are answer spans in the input code snippet with respect to the input query. State-of-the-art BERT-style neural architectures were trained using the NlCodeQA data. Preliminary experimental results show that the proposed model achieves the exact match accuracy of 86.30%. <br> The proposed use of natural language query and neural models for query understanding will help increase the productivity of software developers and pave the way for designing machine learning based code analysis tools that can complement the existing code analysis systems for complex code queries that are either hard or expensive to represent using a formal query language. <br> <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjAzZTBhYzQtYmVjZC00MjEyLTlkZDUtYTNmMTE4OGQ0MTdh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjAzZTBhYzQtYmVjZC00MjEyLTlkZDUtYTNmMTE4OGQ0MTdh%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20220428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220428T120000Z UID:e8adbb2d85a7ac3e47c86770c3f3399d-273 DTSTAMP:19700101T120010Z DESCRIPTION:A Context-Aware Neural Approach for Explainable Citation Link Prediction URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/273/a-context-aware-neural-approach-for-explainable-citation-link-prediction/ SUMMARY:Citations have become an integral part of scientific publications. They play a crucial role in supporting authors claims throughout a scientific paper. However, citing related work is a challenging and laborious task, especially for novice researchers who are not much familiar with the literature and have little or no experience in writing citation text. In this work, we study the task of Citation Link Prediction and propose a novel neural architecture called ExCite, that predicts the existence of a citation link between a pair of scientific documents within a given context. More importantly, it also generates the corresponding citation text at the same time. For this purpose, ExCite leverages diverse role-based views of the documents to learn robust document representations. The proposed model achieves state-of-the-art performance on both citation link prediction and citation text generation subtasks. We performed an extensive set of experiments to show the effectiveness of each module in the proposed neural architecture and evaluated our explanations using a wide range of state-of-the-art automatic evaluation metrics. By performing qualitative and quantitative analyses, we showed that ExCite is capable of generating high-quality citation text that is highly coherent with the citation context. <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzQxMmQ2OTktNGVlNi00NDI0LTgwYTQtNjc5ZmNlMWJjNDhk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzQxMmQ2OTktNGVlNi00NDI0LTgwYTQtNjc5ZmNlMWJjNDhk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20220428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220428T120000Z UID:a1130a5b5ffbc8d2d7009649d4bc661a-274 DTSTAMP:19700101T120009Z DESCRIPTION:Boolean Functional Synthesis using Gated Continuous Logic Networks URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/274/boolean-functional-synthesis-using-gated-continuous-logic-networks/ SUMMARY:Boolean Functional Synthesis (BFS) is a well-known challenging problem in the domain of automated program synthesis from logical specifications. This problem aims to synthesize a Boolean function that is correct-by-construction with respect to the declared specification; this specification symbolically relates the inputs and outputs of the function to be synthesized. Since Boolean functions are the basic building blocks of modern digital systems, BFS has applications in a wide range of areas, including QBF-SAT solving, circuit repair and debugging. This has motivated the community to develop practically efficient algorithms for synthesizing compact Boolean functions, which is a non-trivial endeavor. However, to the best of our knowledge, current techniques are unable to specify a bound on the Boolean function size during synthesis. Specifying a bound on the size of the formula offers flexibility in synthesizing minimal-sized Boolean functions. <br> Learning Boolean functions from logical specifications using neural networks is a difficult problem as it requires the network to represent Boolean functions. Boolean functions are discrete functions and consequently, non-differentiable. Thus, learning a Boolean function directly using traditional neural networks is not possible. Recently Ryan et al proposed the Gated Continuous Logic Network (GCLN) model that builds on Fuzzy Logic to represent Boolean and linear integer operator, in the context of learning invariants for programs. In this work, we investigate the use of the GCLN model to synthesize solutions to the BFS problem. Our model lets us bound the number of clauses used in the synthesized Boolean function. <br> We implement this approach in our tool BNSynth (for Bounded Neural Synthesis), that also uses sampling and counterexample guided techniques to synthesize Boolean functions. We validate our hypothesis that this system can learn smaller functions as compared to a state-of-the-art tool, over custom benchmarks. We observe a 2.4X average improvement in formula size, for these benchmarks. This empirically proves that our system is capable of synthesizing smaller Boolean functions as compared to the state of the art. <br> <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzJlZDc0YTMtYTRmYS00N2Y0LWE5YWQtYzI1YjY2OGI2ODk1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22e2a9daf6-b4d3-4abe-ae30-2ccb517ed18d%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MzJlZDc0YTMtYTRmYS00N2Y0LWE5YWQtYzI1YjY2OGI2ODk1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22e2a9daf6-b4d3-4abe-ae30-2ccb517ed18d%22%7d</a> DTSTART:20220428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220506T120000Z UID:79fa351ee43b1e23a89dcc8ae28c66e0-275 DTSTAMP:19700101T120011Z DESCRIPTION:New techniques and results for Support Recovery in Mixture Models URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/275/new-techniques-and-results-for-support-recovery-in-mixture-models/ SUMMARY:Mixture models with high dimensional parameter vectors are widely used to fit complex multimodal datasets as they allow representation of latent sub-populations within the overall population. The primary difficulty in learning mixture models is that the observed data does not identify the subpopulation to which an individual observation belongs. <br> We study the problem of support recovery in mixture models parameterized by sparse vectors i.e. our goal is to recover the set of non-zero indices of each of the unknown vectors. We present a very generic framework (including a novel tensor-based algorithm) for support recovery by using estimates of the number of unknown vectors having non-zero entries in small subsets of indices. We apply this framework by showing a variety of techniques to estimate the aforementioned quantities in different mixture models. Our results for support recovery are quite general, namely they are applicable to 1) Mixtures of many different canonical distributions including Uniform, Poisson, Laplace, Gaussians, etc. 2) Mixtures of linear regressions and linear classifiers. <br> Based on joint works (https://arxiv.org/abs/2202.11940, https://arxiv.org/abs/2106.05951) with Arya Mazumdar and Venkata Gandikota <br> ----------- <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220506T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220511T120000Z UID:4b96e11b3673af607bcb7b678737ff14-276 DTSTAMP:19700101T120014Z DESCRIPTION:Contextual Concurrency Control URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/276/contextual-concurrency-control/ SUMMARY:Kernel synchronization primitives are of paramount importance to achieving good performance and scalability for applications. However, they are usually invisible and out of the reach of application developers. Instead, kernel developers and synchronization experts make all the decisions regarding kernel lock design. <br> In this talk, I will propose a paradigm, called contextual concurrency control (C3), that enables applications to tune concurrency control in the kernel. C3 allows developers to change the behavior and parameters of kernel locks, to switch between different lock implementations and to dynamically profile one or multiple locks for a specific scenario of interest. This approach opens up a plethora of opportunities to fine-tune concurrency control mechanisms on the fly. DTSTART:20220511T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220512T120000Z UID:211cac5675f6f1390efa6cf6e4c62ed9-277 DTSTAMP:19700101T120021Z DESCRIPTION:Finding Adversarially Robust Representations URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/277/finding-adversarially-robust-representations/ SUMMARY:Adversarial robustness measures the susceptibility of a machine learning algorithm to small perturbations made to the input either at test time or at training time. Our current theoretical understanding of adversarial robustness is limited, and has mostly focused on supervised learning tasks. In this talk, I will consider a natural extension of Principal Component Analysis (PCA) where the goal is to find a low dimensional subspace to represent the given data with minimum projection error, and that is in addition robust to small perturbations. Unlike PCA which is solvable in polynomial time, this formulation is computationally intractable to optimize as it generalizes a well-studied sparse PCA problem. I will describe an efficient algorithm that find approximately optimal solutions and show how this can be used as a robust subroutine for many downstream learning tasks, including training more certifiably robust neural networks. Based on joint works with Pranjal Awasthi, Xue Chen, Vaggos Chatziafratis, Himanshu Jain and Ankit Singh Rawat. ----------- Microsoft Teams Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220512T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220511T120000Z UID:724f4b68fbae88cc530f4242183ac820-278 DTSTAMP:19700101T120014Z DESCRIPTION:Contextual Concurrency Control URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/278/contextual-concurrency-control/ SUMMARY:Kernel synchronization primitives are of paramount importance to achieving good performance and scalability for applications. However, they are usually invisible and out of the reach of application developers. Instead, kernel developers and synchronization experts make all the decisions regarding kernel lock design. In this talk, I will propose a paradigm, called contextual concurrency control (C3), that enables applications to tune concurrency control in the kernel. C3 allows developers to change the behavior and parameters of kernel locks, to switch between different lock implementations and to dynamically profile one or multiple locks for a specific scenario of interest. This approach opens up a plethora of opportunities to fine-tune concurrency control mechanisms on the fly. DTSTART:20220511T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220519T120000Z UID:282f41ec5522d263dcfbc0451781122d-279 DTSTAMP:19700101T120016Z DESCRIPTION:Automating Distributed Heterogeneous Computing for Python Programmers URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/279/automating-distributed-heterogeneous-computing-for-python-programmers/ SUMMARY:Multiple simultaneous disruptions are currently under way in both hardware and software, as we consider the implications for future HPC systems. In hardware, “extreme heterogeneity” has become critical to sustaining cost and performance improvements after Moores Law, but poses significant productivity challenges for developers. In software, the rise of large-scale data science and AI applications is being driven by domain experts from diverse backgrounds who demand the programmability that they have come to expect from high-level languages like Python. While current foundations for compiler and runtime technologies have served us well for many decades, we now see signs of their limitations in the face of these disruptions. This talk makes a case for new approaches to enable productivity and programmability of future HPC systems for domain scientists, and discusses preliminary results obtained in response to the challenge of automating distributed heterogeneous computing for Python programmers. DTSTART:20220519T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220520T120000Z UID:85295fc84184068b4020d53486b7831e-280 DTSTAMP:19700101T120011Z DESCRIPTION:Interventional Complexity of Learning Causal Graphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/280/interventional-complexity-of-learning-causal-graphs/ SUMMARY:A well-studied challenge that arises in the structure learning problem of causal directed acyclic graphs (DAG) is that using observational data, one can only learn the graph up to a "Markov equivalence class" (MEC). The remaining undirected edges have to be oriented using interventions, which can be very expensive to perform in applications. Thus, the problem of minimizing the number of interventions needed to fully orient the MEC has received a lot of recent attention. In this talk, I will describe a new universal lower bound on the number of single-node interventions that any algorithm (whether active or passive) would need to perform in order to orient a given MEC. I will then discuss the tightness of this lower bound and compare it with previously known lower bounds. I will also describe the notion of CBSP orderings, which are topological orderings of DAGs without v-structures, and underly the main technical idea for proving our lower bound. <br> This is based on joint work with Piyush Srivastava (TIFR) and Gaurav Sinha (Adobe Research). Paper link - https://arxiv.org/abs/2111.05070 <br>----------- <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a><br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ DTSTART:20220520T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220520T120000Z UID:f9803be0de33331bb840c8d9c4568231-281 DTSTAMP:19700101T120016Z DESCRIPTION:Recent Research in Machine Perception at Google URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/281/recent-research-in-machine-perception-at-google/ SUMMARY:In this talk I will present some recent progress in machine perception at Google Research, both on fundamental research problems and on the tech powering several popular Google products. I will give a historical retrospective on some important research problems, a snapshot of some relevant current work and observations about the future. And on the applied side, I will provide some background on how we build systems at Google that interpret, reason about and transform sensory data -- and why this aspect of AI is increasingly critical across so many products. DTSTART:20220520T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220519T120000Z UID:fcaed2a7930fa0d4485ae5436d15ae81-282 DTSTAMP:19700101T120016Z DESCRIPTION:Automating Distributed Heterogeneous Computing for Python Programmers URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/282/automating-distributed-heterogeneous-computing-for-python-programmers/ SUMMARY:Multiple simultaneous disruptions are currently under way in both hardware and software, as we consider the implications for future HPC systems. In hardware, “extreme heterogeneity” has become critical to sustaining cost and performance improvements after Moores Law, but poses significant productivity challenges for developers. In software, the rise of large-scale data science and AI applications is being driven by domain experts from diverse backgrounds who demand the programmability that they have come to expect from high-level languages like Python. While current foundations for compiler and runtime technologies have served us well for many decades, we now see signs of their limitations in the face of these disruptions. This talk makes a case for new approaches to enable productivity and programmability of future HPC systems for domain scientists, and discusses preliminary results obtained in response to the challenge of automating distributed heterogeneous computing for Python programmers. DTSTART:20220519T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220520T120000Z UID:55d8d8b368de6d37f23b5bcb10f2e5b3-283 DTSTAMP:19700101T120016Z DESCRIPTION:Recent Research in Machine Perception at Google URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/283/recent-research-in-machine-perception-at-google/ SUMMARY:In this talk I will present some recent progress in machine perception at Google Research, both on fundamental research problems and on the tech powering several popular Google products. I will give a historical retrospective on some important research problems, a snapshot of some relevant current work and observations about the future. And on the applied side, I will provide some background on how we build systems at Google that interpret, reason about and transform sensory data -- and why this aspect of AI is increasingly critical across so many products. DTSTART:20220520T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220601T120000Z UID:da2a20e4af9b0a1fcbc043e0cd1b8451-284 DTSTAMP:19700101T120016Z DESCRIPTION:Enabling Serverless Research with the vHive Open-Source Ecosystem URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/284/enabling-serverless-research-with-the-vhive-open-source-ecosystem/ SUMMARY:Serverless (also known as Function-as-a-Service) has emerged as the next dominant cloud architecture. In serverless, developers structure their cloud application as a collection of stateless functions that execute on-demand. Despite accepted benefits to both developers and cloud providers, the serverless model brings new challenges calling for cross-layer characterization and optimization. Alas, today's state-of-the-art serverless stacks are propriety to each individual cloud provider, which impedes research in this space. <br> In this talk, I will overview the serverless model and introduce the vHive open-source ecosystem developed at the Edinburgh Architecture and Systems (EASE) Lab. vHive includes a complete serverless stack, a suite of serverless workloads and robust tools for performance analysis and debug, which together enable meaningful serverless research at any scale. I will describe two vHive-enabled projects targeting performance analysis and optimization across the system stack that underscore both challenges and opportunities in today's serverless deployments. <br> <br> Talk URL: <a href="https://tinyurl.com/2u8yaxsy">https://tinyurl.com/2u8yaxsy</a> DTSTART:20220601T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220603T120000Z UID:4295a0a9dc9ce5f6a8ee57f8dad565cc-285 DTSTAMP:19700101T120011Z DESCRIPTION:Perfect matchings and Quantum Physics URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/285/perfect-matchings-and-quantum-physics/ SUMMARY:In 2017, Krenn reported that certain problems related to the perfect matchings and colourings of graphs emerge out of studying the constructability of general quantum states using modern photonic technologies. He realized that if we can prove that the weighted matching index of a graph, a parameter defined in terms of perfect matchings and colourings of the graph is at most 2, that could lead to exciting insights on the potential of resources of quantum inference. Motivated by this, he conjectured that the weighted matching index of any graph is at most 2. The first result on this conjecture was by Bogdanov, who proved that the (unweighted) matching index of graphs (non-isomorphic to K_4) is at most 2, thus classifying graphs non-isomorphic to K_4 into Type 0, Type 1 and Type 2. By definition, the weighted matching index of Type 0 graphs is 0. We give a structural characterization for Type 2 graphs, using which we settle Krenns conjecture for Type 2 graphs. Using this characterization, we provide a simple O(|V||E|) time algorithm to find the unweighted matching index of any graph. In view of our work, Krenns conjecture remains to be proved only for Type 1 graphs. We give upper bounds for the weighted matching index in terms of connectivity parameters for such graphs. Using these bounds, for a slightly simplified version, we settle Krenns conjecture for the class of graphs with vertex connectivity at most 2 and the class of graphs with maximum degree at most 4. For our full paper, see paper link (https://arxiv.org/abs/2202.05562). --------- For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/?talk=20220603_RishikeshGajjala DTSTART:20220603T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220614T120000Z UID:7a5dc6323dd21af29cbe1912b3e1a887-286 DTSTAMP:19700101T120016Z DESCRIPTION:Leveraging AI & HPC to Enable Semiconductor Manufacturing in the Angstrom-era URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/286/leveraging-ai-hpc-to-enable-semiconductor-manufacturing-in-the-angstrom-era/ SUMMARY:Semiconductor manufacturing is approaching the Angstrom-era with innovations such as gate all around transistors, and chip to chip integration technologies enabling the continuation of Moores law. This talk will highlight some of the challenges that these advanced technologies pose to manufacturing semiconductors, and will cover how modern AI & HPC technologies are being leveraged to address these challenges to enable high-volume manufacturing. We will also give a peek into some of the solutions that KLA is pioneering in this space. DTSTART:20220614T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220603T120000Z UID:88f5a08a3c0ac64480eb247bad114d55-288 DTSTAMP:19700101T120015Z DESCRIPTION:Wafer-Scale processing for AI and HPC URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/288/wafer-scale-processing-for-ai-and-hpc/ SUMMARY:As ML algorithms and network architectures evolve rapidly to improve performance for variety of applications, some trends are very clear: larger models are often highly effective, compute and memory requirements are growing exponentially, scale-out based approaches are yielding sub-linear results, deep expertise is needed to achieve high utilization on scale-out. These trends motivate for Wafer-Scale chip with distributed memory architecture and a domain specific ISA for ML applications. It turns out that this architecture can also offer compelling performance advantage on some HPC applications. This talk covers some of the key details of Cerebras Wafer-Scale Engine (WSE) based accelerator and how it accelerates AI and HPC applications. DTSTART:20220603T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220610T120000Z UID:1f93090caaa0f047e60589da68989f8d-291 DTSTAMP:19700101T120011Z DESCRIPTION:Determinant Maximization: Approximation and Estimation Algorithms URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/291/determinant-maximization-approximation-and-estimation-algorithms/ SUMMARY:In the determinant maximization problem, given a collection of vectors, we aim to pick a subset to maximize the determinant of a natural matrix associated with these vectors. The abstract problem captures problems in multiple areas including machine learning, statistics, convex geometry, Nash social welfare problem from algorithmic game theory and network design problems. We will survey the known results and techniques for the problem. The results vary from arbitrary good approximations to only estimation algorithms. The techniques used in these works vary from geometry of polynomials, sparse solutions to convex programming solutions to matroid intersection algorithms. <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar DTSTART:20220610T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220622T120000Z UID:de6cf0ed392f45a2b0187c45d4360a10-292 DTSTAMP:19700101T120014Z DESCRIPTION:Explainable and Efficient Neural models for Natural Language to Bash Command Translation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/292/explainable-and-efficient-neural-models-for-natural-language-to-bash-command-translation/ SUMMARY:One of the key goals of Natural Language Processing is to make computers understand natural language. Semantic Parsing has been one of the driving tasks for Natural Language Understanding. It is formally defined as the task of generating meaning representation from natural language input. In this work, we focus on using the Bash command as the meaning representation. Bash is a Unix command language used for interacting with the Operating System. Recent works on natural language to Bash command translation have made significant advances on this problem. The best performing solutions employ a neural network architecture called the Transformer. In this work, we explore the aspects of explainability and efficiency for this task and use the Transformer as one of the baselines for comparing the proposed approaches. <br> In the first part, we utilize documentation data from Linux manual pages and the Abstract Syntax Tree for Bash to generate explanations for the translated Bash command. We propose a novel architecture that incorporates tree structure information in the Transformer and provides explanations for its predictions via alignment matrices between user invocation and manual page text. We find that the proposed method performs on par with the Transformer performance. Our method performs better than fine-tuned T5, a Transformer-based neural model pre-trained on a large amount of text data in a self-supervised manner. <br> In the second part, we use the problems inherent synchronous structure and propose the Segmented Invocation Transformer (SIT) that utilizes the information from the constituency parse tree of the natural language invocation. Our method is motivated by the alignment between segments in the natural language text and Bash command components. By utilizing this structure, the proposed method outperforms the state-of-the-art approach while achieving a 1.8x improvement in the inference time (as measured on a CPU) and a 5x reduction in model parameters. We also conduct an attribution analysis using Integrated Gradients to empirically confirm the identified structure and the ability of SIT to capture it. <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_MjZhOWViYzMtZmQ5MC00NzMwLWI3MjktODlhOGU4YjkxZGYz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_MjZhOWViYzMtZmQ5MC00NzMwLWI3MjktODlhOGU4YjkxZGYz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d</a> DTSTART:20220622T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220617T120000Z UID:eec8218c1d92055ae6295b3e08751ad8-293 DTSTAMP:19700101T120015Z DESCRIPTION:Why should we learn logic, and why should we learn logic? URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/293/why-should-we-learn-logic-and-why-should-we-learn-logic/ SUMMARY:Computer science curricula do not usually advertise the fact that computers, programming and models of computation all arose from trying to answer fundamental questions in logic. But is it only a matter of historical pride? Could learning logic actually be useful for those who actually do things (and not just theorize)? Even if that is true, should it not be sufficient for a few to write formulas and such, while most of us build "apps"? These are reasonable questions for a computer science student to ask, and the talk is an attempt to engage in a discussion with her. DTSTART:20220617T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220616T120000Z UID:ed08a1adce7069bf211f08a1443ed675-294 DTSTAMP:19700101T120021Z DESCRIPTION:Hypergraph expansion, CSPs, and algorithmic decoding of epsilon-balanced codes URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/294/hypergraph-expansion-csps-and-algorithmic-decoding-of-epsilon-balanced-codes/ SUMMARY:We will discuss some new notions of hypergraph expansion, which can be exploited by spectral algorithms, as well as ones based on semidefinite programming hierarchies. These properties lead to new structural characterizations and algorithmic regularity lemmas for hypergraphs, as well as new decoding algorithms for codes based on bias-reduction via direct-sum, such as the breakthrough construction of epsilon-balanced codes by Ta-Shma. (Based on joint work with Vedat Levi Alev, Fernando Granha Jeronimo, Dylan Quintana, and Shashank Srivastava) <br> <br> For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ <br> <br> Microsoft Teams Link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20220616T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220621T120000Z UID:f57e8d34fd4e272d0674728198421553-295 DTSTAMP:19700101T120010Z DESCRIPTION:HYDRA: A Dynamic Approach to Database Regeneration URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/295/hydra-a-dynamic-approach-to-database-regeneration/ SUMMARY:Database software vendors often need to generate synthetic databases for a variety of applications, including (a) Testing database engines and applications, (b) Data masking, (c) Benchmarking, (d) Creating what-if scenarios, and (e) Assessing performance impacts of planned engine upgrades. The synthetic databases are targeted toward capturing the desired schematic properties (e.g.~keys, referential constraints, functional dependencies, domain constraints), as well as the statistical data profiles (e.g., value distributions, column correlations, data skew, output volumes) hosted on these schemas. <br> Several data generation frameworks have been proposed in the last two decades. It started from the ab-initio generation tools that use standard mathematical distributions and do not depend on the client databases or query workloads. Subsequently, tools that generate data using column distributions became prominent. However, none of these mechanisms could mimic the customer query-processing environments satisfactorily. The more contemporary school of thought is generating workload-aware data that uses query execution plans from the customer workloads as input and guarantees volumetric similarity. That is, the intermediate row-cardinalities obtained at the client and vendor sites are very similar when matching query plans are executed. This similarity helps to preserve the multi-dimensional layout and flow of the data, a prerequisite for achieving similar performance on the clients workload. However, even in this category, the existing frameworks are crippled by one or more of the following limitations: (a) inability to provide a comprehensive algorithm to handle the queries based on core relational algebra operators, namely, select, project, and join; (b) inability to scale to big data volumes; (c) inability to scale to large input workloads; and (d) poor accuracy on unseen queries. <br> In this work, motivated by the above lacuna, we present HYDRA, a data regeneration tool that materially addresses the above challenges by adding functionality, dynamism, scale, and robustness. Firstly, the extended workload coverage is obtained by providing a comprehensive solution to support queries based on select-project-join relational algebra operators. Specifically, the constraints are modeled using a linear feasibility problem, in which each variable represents the volume of a region of the data space. These regions are computed using a scheme of partitioning strategies. For example, to encode the filter constraints, our region-partitioning approach divides the data space into the provably minimum number of regions, thereby reducing the existing solutions complexity by many orders of magnitude. Our projection subspace division and projection isolation strategies address the critical challenges in incorporating projection-inclusive constraints. By modeling referential constraints over denormalized equivalents of the tables, Hydra delivers a comprehensive solution that also additionally handles join constraints. <br> Secondly, a unique feature of our data regeneration approach is that it delivers a database summary as the output rather than the static data itself. This summary is of negligible size and depends only on the query workload and not on the database scale. It can be used for dynamically generating data during query execution. Therefore, the enormous time and space overheads incurred by prior techniques in generating and storing the data before initiating analysis are eliminated. Specifically, the summaries for complex Big Data client scenarios comprising over a hundred queries are constructed within just a few minutes, requiring only a few 100 KBs of storage. We have evaluated the proposed ideas using both synthetic benchmarks such as TPC-DS and real-world benchmarks based on Census and IMDB databases. <br> Thirdly, to improve accuracy towards unseen queries, Hydra additionally exploits metadata statistics maintained by the database engine. Specifically, it adds an objective function to the linear program to pick a solution with improved inter-region tuple distribution. Further, a uniform distribution of tuples within regions is generated to get a spread of values. In a nutshell, these techniques facilitate careful selection of a desirable database from the candidate synthetic databases and also provide metadata compliance. <br> Lastly, as a proof of concept, the Hydra framework has been prototyped in a Java based-tool that provides a visual and interactive demonstration of the data regeneration pipeline. The tool has been warmly received by both academic and industrial communities. DTSTART:20220621T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220628T120000Z UID:e5a732e76581febc5598265416564215-296 DTSTAMP:19700101T120015Z DESCRIPTION:Improved Algorithms for Variants of Bin Packing and Knapsack. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/296/improved-algorithms-for-variants-of-bin-packing-and-knapsack/ SUMMARY:We study variants of two classical optimization problems: Bin Packing and Knapsack. Both bin packing and knapsack fall under the regime of & Packing and Covering Problems&quot;. In bin packing, we are given a set of input items, each with an associated size, and the objective is to pack these into the minimum number of unit capacity bins. On the other hand, in the knapsack problem, each item has an additional profit associated with it. The objective is to find a maximum profitable subset that can be packed into a unit capacity knapsack. Both bin packing and knapsack find numerous applications; however, both turn out to be NP-Hard. Hence, it is natural to seek approximation algorithms for these problems. Lawler settled the knapsack problem by giving an FPTAS, whereas the progressive works of de la Vega and Lueker, Karmarkar and Karp, and Rothvoss have more-or-less settled the bin packing problem. However, many variants of these problems (e.g., multidimensional, geometric, stochastic) also find wide applicability, but havent been settled. We make progress on this front by providing new and improved algorithms for several such variants. <br> First, we study bin packing under the i.i.d. model, where item sizes are sampled independently and identically from a distribution in (0,1]. Both the distribution and the total number of items are unknown. The items arrive one by one, and their sizes are revealed upon their arrival, and they must be packed immediately and irrevocably in bins of unit size. We provide a simple meta-algorithm that takes an offline alpha-asymptotic approximation algorithm and provides a polynomial-time (alpha+epsilon)-competitive algorithm for online bin packing under the i.i.d. model, where epsilon&gt;0 is a small constant. Using the AFPTAS for offline bin packing, we thus provide a linear time (1+epsilon)-competitive algorithm for online bin packing under the i.i.d. model, thus settling the problem. <br> Then we study a well-known geometric generalization of the knapsack problem, the 3D Knapsack problem. In this problem, the items are cuboids in three dimensions, and the knapsack is a unit cube. The objective is to pack a maximum profitable subset of the input set in a non-overlapping, axis-parallel manner inside the knapsack. Depending on whether rotations around axes (by ninety degrees) are allowed or not, we obtain two variants. [DHJTT 07] gave a (7+epsilon) (resp. (5+epsilon)) approximation algorithm for the 3D Knapsack problem without rotations (resp. with rotations). Despite the importance of the problem, there has been no improvement in the ratios for fifteen years. First, we give alternate algorithms that achieve the same approximation ratios (7+epsilon, 5+epsilon). These algorithms and their analyses are far simpler. Then, for the case when rotations are allowed, we give an improved (31/7+epsilon) approximation algorithm in the general setting, and a 2.78 approximation algorithm in the important special case where each item has a profit equal to its volume. <br> We also introduce and study a generalization of the knapsack problem with geometric and vector constraints. The input is a set of rectangular items, each with an associated profit and d nonnegative weights (dD vector), and a square knapsack. The goal is to find a non-overlapping, axis-parallel packing of a subset of items into the given knapsack such that the vector constraints are not violated, i.e., the sum of weights of all the packed items in any of the d dimensions does not exceed one. Two variants are defined: rotations allowed by 90 degrees and rotations not allowed. We give (2+epsilon)-approximation algorithms for both variants. <br> Finally, we consider the problem of packing dD hypercubes into a knapsack defined by the region [0,1]^d. Each hypercube has an associated profit, and the goal is to find a maximum profitable non-overlapping, axis-parallel packing. We consider two special cases of this problem: (i) cardinality case, where each item has unit profit, (ii) bounded profit-volume ratio case, where the profit-to-volume ratio of each item lies in the range [1,r] for some fixed constant r. We give near-optimal algorithms for both cases. <br> <br> Microsoft teams link:<br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZmVjYTEwZWQtN2RlMy00ZDVmLWJiZTItOGU0MDc3NTRlMjk0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2238682e77-d26b-4d00-8761-a89f372adc36%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZmVjYTEwZWQtN2RlMy00ZDVmLWJiZTItOGU0MDc3NTRlMjk0%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2238682e77-d26b-4d00-8761-a89f372adc36%22%7d</a> DTSTART:20220628T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220629T120000Z UID:13210eae64431aac45fbee44ec633fa8-297 DTSTAMP:19700101T120009Z DESCRIPTION:Inducing constraints in paraphrase generation and consistency in paraphrase detection URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/297/inducing-constraints-in-paraphrase-generation-and-consistency-in-paraphrase-detection/ SUMMARY:Deep learning models typically require a large volume of data. Manual curation of datasets is time-consuming and limited by imagination. As a result, natural language generation (NLG) has been employed to automate the process. However, NLG models are prone to producing degenerate, uninteresting, and often hallucinated outputs [1]. Constrained generation aims to overcome these shortcomings by appending additional information to the generation process. Training data thus generated can help improve the robustness of deep learning models. Therefore, the key research question of the thesis is: “How can we constrain generation models, especially in NLP, to produce meaningful outputs and utilize them for building better classification models?” In the first part, we present two approaches for constraining NLG models via the task of paraphrase generation. Paraphrase generation involves the generation of text that conveys the same meaning as a reference text. Our proposal is the following two strategies: DiPS (Diversity in Paraphrases using Submodularity): The first approach deals with constraining paraphrase generation to ensure diversity, i.e., ensuring that generated text(s) are sufficiently lexically different from each other without compromising on relevance (fidelity). We propose a decoding algorithm for obtaining such diverse texts. We provide a novel formulation of the problem in terms of monotone submodular function maximization, specifically targeted toward the task of paraphrase generation. We demonstrate the effectiveness of our method for data augmentation on multiple tasks such as intent classification and paraphrase recognition. SGCP (Syntax Guided Controlled Paraphraser): The second approach deals with constraining paraphrase generation to ensure syntacticality, i.e., ensuring that the generated text is syntactically coherent with an exemplar sentence. We propose Syntax Guided Controlled Paraphraser (SGCP), an end-to-end framework for syntactic paraphrase generation without compromising relevance (fidelity). The framework uses a sequence-to-sequence model with a Tree-LSTM-based gating mechanism to selectively choose syntactic representations during text generation. This approach performs significantly better than prior works that utilize only limited syntactic information in the exemplar. The second part of the research question pertains to ensuring that the generated output is meaningful. ​​Towards this, we present an approach for paraphrase detection to ascertain that the generated output is semantically coherent with the reference text. Paraphrase Detection is the task of detecting whether or not the two input natural language statements are paraphrases of each other. Fine-tuning pre-trained models such as BERT and RoBERTa on paraphrase datasets have become the go-to approaches for such tasks. However, tasks like paraphrase detection are symmetric - they require the output to be invariant of the order of the inputs. In fine-tuned models for classification, inconsistency is often observed in the predicted labels or confidence scores. We validate this shortcoming and apply a consistency loss function to alleviate inconsistency in symmetric classification. Our results show an improved consistency in predictions for three paraphrase detection datasets without a significant drop in the accuracy scores. While this work addresses the research question via paraphrase generation and detection, the approaches presented here apply broadly to NLP-based deep learning models that require imposing constraints and ensuring consistency. The work on paraphrase generation can be extended to impose new kinds of constraints on generation (for example, sentiment coherence), and the work on paraphrase detection can be applied to ensure consistency in other symmetric classification tasks that use deep learning models (for example, sarcasm interpretation). References: [1] Ehud Reiter. Hallucination in neural nlg. https://ehudreiter.com/2018/11/12/hallucination-in-neural-nlg/, 2018. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGM0NjE1ZTYtNWJhMi00ZjhmLTkxOTMtNjBhMmY1ZDYxMGI3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2258c2a810-9762-463d-90d0-20832b3d1592%22%7d DTSTART:20220629T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220701T120000Z UID:813c46135406f4d06957d7770e0ace72-302 DTSTAMP:19700101T120011Z DESCRIPTION:Online Algorithm for the Minimum Metric Bipartite Matching Problem URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/302/online-algorithm-for-the-minimum-metric-bipartite-matching-problem/ SUMMARY:In the online minimum-metric bipartite matching (OMBM) problem, we are given a set S of server locations. The locations of requests (given by the set R) are revealed one at a time and when a request is revealed, we must immediately and irrevocably match it to a free DTSTART:20220701T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220720T120000Z UID:8e4cab9983eb002af5ae3a1de3bee15a-303 DTSTAMP:19700101T120011Z DESCRIPTION:Moving Fast with High Reliability using Pluggable Types URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/303/moving-fast-with-high-reliability-using-pluggable-types/ SUMMARY:For many real-world applications, software reliability is of critical importance. At the same time, developers need to be able to move fast in developing new features and products. In this talk, I will describe recent work on using pluggable type systems to reduce the tension between these seemingly-conflicting needs. First, I will present NullAway, a novel nullability type system for Java. NullAway improves on previous work by reducing build-time overhead and requiring fewer annotations through carefully-targeted unsoundness. Then, I will describe more recent work on lightweight and modular typestate analysis targeting accumulation properties, a class of typestate properties that can be checked soundly without heavyweight alias analysis. I will present two instantiations of this approach: the Object Construction Checker, a type system to ensure the safe usage of builders and other complex initialization schemes, and the Resource Leak Checker for practical prevention of resource leaks. DTSTART:20220720T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220718T120000Z UID:4e840d61a0fbc917b3bd86e434dd7f17-304 DTSTAMP:19700101T120013Z DESCRIPTION:Operating System Support for Efficient Virtual Memory URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/304/operating-system-support-for-efficient-virtual-memory/ SUMMARY:Online talk URL: <a href=" https://tinyurl.com/46d5f3wx">https://tinyurl.com/46d5f3wx</a> <br> Virtual memory has proven to be an extremely powerful abstraction for its programmability benefits. Unfortunately, virtual memory is becoming a performance bottleneck due to the address translation wall. Modern applications with large memory footprints necessitate frequent page table walks to perform the virtual to physical address translation. Consequently, the hardware spends 30-50% of the total CPU cycles in servicing TLB misses alone. Virtualization and non-uniform memory access (NUMA) architectures further exacerbate this overhead. For example, virtualized systems involve two-dimensional page table walks that require up to 24 memory accesses for each TLB miss, with current 4-level page tables. The address translation performance drops further on NUMA systems, depending on the distance between the CPU and page tables. These overheads will increase in the future, where deeper page tables and multi-tiered memory systems will enable even larger applications. Virtual memory, therefore, is showing its age in the era of data-centric computing. <br> This thesis investigates the role of an operating system (OS) and hypervisor in improving the address translation performance. First, we focus on huge pages that can significantly reduce the frequency and cost of TLB misses. Huge pages are widely available in modern systems e.g., x86 architecture supports 2MB and 1GB huge pages, in addition to regular 4KB pages. While huge pages are great in theory, real-world OSs have often delivered disappointing performance while using them. This is because memory management of huge pages is fraught with multiple challenges. We propose several enhancements in OS-level policies and mechanisms to make huge pages beneficial, even under multi-dimensional constraints such as latency, capacity, and fairness. <br> Second, we investigate the effect of NUMA on address translation performance. NUMA architectures mandate careful data placement to hide the effect of variable memory access latency from applications. Several decades of research on NUMA systems have optimized access to user-level application data. However, prior research has ignored the access performance of kernel data, including page tables, due to their small memory footprint. We argue that it is time to revisit page table management for NUMA-like systems. <br> <br> The core contributions of this thesis include four systems: Illuminator, HawkEye, Trident, and vMitosis, as summarized below: <br> Illuminator: We first expose some subtle implications of external memory fragmentation on huge pages. We show that despite proactive measures employed in the memory management subsystem of Linux, unmovable kernel objects (e.g., inodes, page tables, etc.) can deny huge pages to user applications. In a long-running system, unmovable objects fragment physical memory, often permanently, and cause high de-fragmentation overheads. Over time, their effects manifest in performance regressions, OS jitter, and latency spikes. Illuminator effectively clusters kernel objects in a subset of physical memory regions and makes huge page allocations feasible even under heavily fragmented scenarios.. <br> HawkEye: In this work, we deal with OS-based huge page management policies that need to balance complex trade-offs between TLB coverage, memory bloat, latency, and the number of page faults. In addition, we consider performance and fairness issues that appear under fragmentation when memory contiguity is limited. In HawkEye, we propose asynchronous page pre-zeroing to simultaneously optimize for low latency and few page faults. We propose automatic bloat recovery to effectively deal with the trade-offs between TLB coverage and memory bloat at runtime. HawkEye addresses the performance and fairness challenges by allocating huge pages based on their estimated profitability. <br> <br> Trident: Illuminator and HawkEye try to extract maximum benefits from 2MB huge pages. However, recent findings have shown that even after employing 2MB pages, more than 20% of the total CPU cycles are wasted in handling TLB misses for data center applications. We address this problem using 1GB huge pages that provide up to 1TB per-core TLB coverage on modern systems. Leveraging insights from our earlier work, we propose a multi-level huge page framework called Trident that judiciously allocates 1GB, 2MB, and 4KB pages as deemed suitable at runtime. <br> vMitosis: In this work, we focus on the effect of NUMA on address translation in virtualized servers. We show that page table walks often involve remote memory accesses on NUMA systems that can slow down large memory applications by more than 3x. Interestingly, the slow down observed due to remote page table accesses can even outweigh that of accessing remote data, even though page tables consume less than 1% memory of overall application footprint. vMitosis mitigates the effect of NUMA on page table walks by enabling each core to handle TLB misses from its local socket. We achieve this by judiciously migrating and replicating page tables across NUMA sockets. <br> <br> Overall, with this thesis, we show that adequate OS and hypervisor support can help virtual memory thrive even in the era of data-centric computing. We have implemented our proposed systems in the Linux OS kernel and KVM hypervisor. Our optimizations are transparent to the users, and using them does not require any hardware or application modifications. DTSTART:20220718T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220715T120000Z UID:01571d4ec8667a411bf1c62b345a0c47-305 DTSTAMP:19700101T120016Z DESCRIPTION:Scheduling to minimize the Age of Information URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/305/scheduling-to-minimize-the-age-of-information/ SUMMARY:Traditional metrics of interest in scheduling literature are flow-time, completion-time, makespan etc. For modern applications e.g. IoT, smart cars, gaming, etc. information timeliness is essential and one measure that captures it well is the Age of Information, which counts how stale the information is. In this talk, we will consider the online scheduling problem of minimizing the Age of Information for the most general system model and discuss some initial progress in terms of deriving online algorithms with constant competitive ratios. For more details about the seminar please visit the website at https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft team link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d DTSTART:20220715T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220713T120000Z UID:388994edafb3d5a74c932a4c7e1c37f0-306 DTSTAMP:19700101T120011Z DESCRIPTION:Bridging control theoretic approaches with methods in formal verification via Induction URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/306/bridging-control-theoretic-approaches-with-methods-in-formal-verification-via-induction/ SUMMARY:Inductive invariants have been widely studied in the context of program verification. In many cases where one fails to show a property is inductive, one instead tries to use different notions of induction (such as double, mutual, $k$, and strong), and then tries to show the property is inductive according to these notions. Analogously, the notion of barrier certificates act as inductive guarantees of safety for discrete-time continuous space systems. Barrier certificates are functions over the state space of the system such that they are non-increasing as the system evolves. The search for these functions typically rely on first deciding on a fixed template (e.g. a fixed degree polynomial) and then solving a semidefinite program to search for such a function. When such a function cannot be found, one typically changes the template (by increasing the degree of the polynomial) and tries again. Inspired by the idea of $k$-induction we considered notions of $k$-inductive barrier certificates and showed that for a fixed degree one may still find $k$-inductive barrier certificates when standard barrier certificates could not be found. As future research, a few natural follow up questions arise: 1) What are the different notions of barrier certificates for the notions of induction? 2) In many cases where one is not able to show a property to be inductive on a program, one may add more details or change the program in some way to help show the inductive property, what is the analogous idea in the context of dynamical systems? Lastly, I shall consider the question in the other direction: What are some of the other areas where control theory may inform approaches to formal verification? DTSTART:20220713T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220728T120000Z UID:d110c4157ccdcb9944c7d36c557682e7-307 DTSTAMP:19700101T120011Z DESCRIPTION:Performance Characterization and Optimizations of Traditional ML Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/307/performance-characterization-and-optimizations-of-traditional-ml-applications/ SUMMARY:In recent years, Deep Learning based methods have attracted a lot of attention and research – both from statistics and systems. These traditional algorithms are easily explainable and are pretty fast for smaller and medium-size datasets. However, in large organizations, massive datasets spanning a couple of million sample points are not rare. A lot of research has been done to design or adapt these traditional algorithms for such massive datasets. However, we find an apparent lack of a detailed systems-based study for these algorithms in the context of huge datasets. In this work, we study the systems behavior and bottlenecks for these algorithms in the context of huge training datasets. As part of our work, we start with a performance characterization study, identify critical performance bottlenecks experienced by these applications, and then measure the reduction in performance stalls along with apparent benefits in terms of speedup after applying some of the well-known optimizations at the levels of caches, main memory, and computation. More specifically, we apply optimizations such as (i) software prefetching to improve cache performance and (ii) data layout and computation reordering optimizations to improve locality in DRAM accesses and show the performance benefits they can bring in these applications. Last, we evaluate the sensitivity of predictions and the improvement in performance when the computations on precise (float/double) inner variables are interpreted as relatively low-cost integer operations. These optimizations are implemented as modification on the well-known scikit-learn library. We evaluate the impact of the proposed optimizations using a combination of simulation and execution on real system and performance measurement. Our optimizations result in performance benefits varying from 5% -- 27% on different ML applications. This is an online defense. The Teams meeting link for this is: https://tinyurl.com/HarshKumarDefense DTSTART:20220728T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220725T120000Z UID:0a59d0fb26ebe451e0e778782820b9b5-308 DTSTAMP:19700101T120010Z DESCRIPTION:On symmetries of and equivalence tests for two polynomial families and a circuit class URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/308/on-symmetries-of-and-equivalence-tests-for-two-polynomial-families-and-a-circuit-class/ SUMMARY:Two n-variate polynomials f and g in F[x1,...,xn] are said to be equivalent over the field F if there exists an invertible matrix A over F such that f = g(Ax), where x = (x1...xn). The problem of testing whether f is equivalent to a polynomial g coming from a polynomial family G (or computed by a circuit C in a circuit class D) is called the equivalence test (in short, ET) for G (respectively, D). In this thesis, we study equivalence tests for the determinant polynomial family and the class of regular read-once arithmetic formulas (ROFs). We also study some structural and algorithmic properties related to the symmetries of the Nisan-Wigderson design polynomial (in short, NW) and solve an interesting special case of ET for NW. In the first work, we study ET for the determinant (in short, DET) over finite fields and the field of rational numbers denoted Q. A randomized polynomial time DET over the field of complex numbers was given by Kayal in [Kay12]. But DET over finite fields and over Q were not known. We give the first randomized polynomial-time DET over finite fields and also give the first DET over Q. The DET over Q takes oracle access to an integer factoring algorithm (IntFact) and if the input polynomial f is equivalent to the n x n determinant over Q, then it outputs a certificate matrix A over Q. This algorithm is randomized and is efficient for bounded values of n. Assuming the generalized Riemann hypothesis, we also show that the problem of integer factoring reduces to DET for quadratic forms (i.e., n = 2 case). If the DET algorithm does not take oracle access to IntFact, then it outputs a certificate matrix over an extension field L of Q, where [L:Q]<=n. This variant of DET is also randomized and is efficient for any value of n. The DET algorithms over finite fields and Q are obtained by decomposing the Lie algebra of f and then invoking known algorithms for the full matrix algebra isomorphism (FMAI) problem over finite fields and Q. FMAI is a well-studied problem in computer algebra. We also give a reduction from FMAI to DET, which is efficient when n is bounded. This is joint work with Ankit Garg, Neeraj Kayal, and Chandan Saha. In the second work, we study ET for read-once arithmetic formulas (ROFs). An ROF is an arithmetic formula where every leaf node is labeled by either a distinct variable or a constant from the underlying field. ROFs are well-studied in the literature. In this work, we give the first randomized polynomial-time ET with oracle access to quadratic form equivalence for certain restricted ROFs, which we call regular ROFs. ET for regular ROFs generalizes the well-known quadratic form equivalence problem over the field of complex numbers and ETs for the classes of sum-product polynomials and ROANFs. ETs for these two classes have been recently studied by Medini & Shpilka (2021). Our ET algorithm uses some crucial properties related to the non-zeroness, the factors, and essential variables of the Hessian determinant of a regular ROF. We study these properties for the Hessian determinant of an arbitrary ROF C by analyzing the structures and coefficients of some nice monomials in the Hessian determinant of C. This is joint work with Chandan Saha and Bhargav Thankey. In the last work, we study some structural and algorithmic properties related to the symmetries of NW and give a special case of ET for NW. In NW, each pair of monomials has very few variables in common. This property of NW has been exploited to give strong lower bounds for different classes of arithmetic circuits. Like NW, other polynomials like the permanent, the determinant, etc., have also been extensively used in lower bound results. But unlike these polynomials, not much is known about NW. In this work, we study some important properties of NW related to its symmetries. A matrix A is said to be a symmetry of NW if NW(Ax) = NW. We show that NW is characterized by its symmetries over the field of complex numbers but not over the fields of real numbers and rational numbers. Using the symmetries of NW, we show that NW is characterized by circuit identities over any field. This result implies a randomized polynomial time circuit testing algorithm for NW - which tests whether some circuit C computes NW- and the flip theorem for NW. We also solve an interesting special case of ET for NW, which we call the block-diagonal permutation scaling ET for NW. This ET uses the symmetries of NW crucially. This is joint work with Chandan Saha. DTSTART:20220725T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220728T120000Z UID:43db3e52a3e14bd544b10667089fb00a-309 DTSTAMP:19700101T120010Z DESCRIPTION:Why we could not prove SETH hardness of CVP for even norms! URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/309/why-we-could-not-prove-seth-hardness-of-cvp-for-even-norms/ SUMMARY:Lattice-based cryptographic schemes have generated much interest in recent years. Their security relies on the computational hardness of problems over geometric objects called lattices. These problems have been used to build advanced cryptographic primitives such as fully homomorphic encryption, and they are believed to be resistant to quantum attacks. Given the recent advancement in quantum technologies, many institutes such as the National Institute of Standards and Technology (NIST) and European Telecommunications Standards Institute (ETSI) have initiated a process for standardization and deployment of lattice-based schemes widely over the next few years. Recently, NIST has announced lattice-based candidates (Kyber and Dilithium) as primary algorithms for implementation. <br> The security of the lattice-based cryptosystem schemes crucially relies on the assumption that the best-known algorithms for the corresponding lattice problems cannot be significantly improved. Understanding the fine-grained hardness of these problems is one way of getting more confidence in these assumptions. <br> In this talk, I will discuss the fine-grained hardness of the lattice problems in different p-norms. Mainly, I will focus on the recent joint work with Divesh Aggarwal. Under a complexity-theoretic assumption, we show that getting any SETH-hardness for the lattice problems in the even norm is impossible by a poly-time reduction from k-SAT to CVP. We also show similar impossibility results for SUBSET-SUM and many other problems. DTSTART:20220728T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220725T120000Z UID:1cde9a4fb26da09c3aa6b246c13c7fde-310 DTSTAMP:19700101T120014Z DESCRIPTION:Comparative Analysis of Topological Structures URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/310/comparative-analysis-of-topological-structures/ SUMMARY:Measuring scientific processes result in a set of scalar functions (scalar fields) which may be related temporally, be part of an ensemble, or unrelated. Overall understanding and visualization of scientific processes require the study of individual fields and, more importantly, the development of methods to compare them meaningfully. In this thesis, we focus on the problem of designing meaningful measures to compare scalar fields by comparing their abstract representations called topological structures. We emphasize on intuitive and practical measures with useful properties and applications. The first part of the thesis deals with comparing a topological structure called the merge tree. We propose two global comparison measures, both based on tree edit distances. The first measure OTED is based on the assumption that merge trees are ordered rooted trees. Upon finding that there is no meaningful way of imposing such an order, we propose a second measure called MTED for comparing unordered rooted trees. We propose intuitive cost models and prove that MTED is a metric. We also provide various applications such as shape comparison, periodicity detection, symmetry detection, temporal summarization, and an analysis of the effects of sub-sampling /smoothing of on the topology of the scalar field. The second part deals with a local comparison measure LMTED for merge trees that supports the comparison of substructures of scalar fields, thus facilitating hierarchical or multi-scale analysis and alleviating some drawbacks of MTED. We propose a dynamic programming algorithm, prove that LMTED is a metric and also provide applications such as symmetry detection in multiple scales, a finer level analysis of sub-sampling effects, an analysis of the effects of topological compression, and feature tracking in time-varying fields. The third part of the thesis deals with comparison of a topological structure called the extremum graph. We provide two comparison measures for extremum graphs based on persistence distortion (PDEG) and Gromov-Wasserstein distance (GWEG). Both persistence distortion and Wasserstein distance are known metrics. We analyze how the underlying metric affects these comparison measures and present various applications such as periodicity detection to facilitate scientific data analysis and visualization. The final part of the thesis introduces a time-varying version of extremum graphs (TVEG) with a simple comparison criterion to identify correspondences between features in successive time steps. We provide applications to tracking features in time-varying scalar fields from computational fluid dynamics. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_YmJlMzljYTYtMTZkMS00MTkxLWEyMDgtNmQ3MzI2NTQzYmQz%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22c747ccaa-ceaa-4197-b4cb-ce2f1d4694da%22%7d DTSTART:20220725T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220805T120000Z UID:13f63a9e969e9f3406ab492cab3a1bc9-311 DTSTAMP:19700101T120011Z DESCRIPTION:Neural Approaches for Natural Language Query Answering over Source Code URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/311/neural-approaches-for-natural-language-query-answering-over-source-code/ SUMMARY:During software development, developers need to ensure that the developed code is bug-free and the best coding practices are followed during the code development process. To guarantee this, the developers require answers to queries about specific aspects of the code relevant to the development. Powerful code-query languages such as CodeQL have been developed for this purpose. Use of such code-query languages, however, requires expertise in writing formal queries. For each separate query, one needs to write several lines in a code-query language. To remedy these problems, we propose to represent each query by a natural language phrase and answer such queries using neural networks. We aim to perform model training such that a single model can answer multiple queries as opposed to writing separate formal queries for each task. Such a model can answer these queries against unseen code. With this motivation, we introduce the novel NlCodeQA dataset. It includes 171,346 labeled examples where each input consists of a natural language query and a code snippet. The labels are answer spans in the input code snippet with respect to the input query. State-of-the-art BERT-style neural architectures were trained using the NlCodeQA data. Preliminary experimental results show that the proposed model achieves the exact match accuracy of 86.30%. The proposed use of natural language query and neural models for query understanding will help increase the productivity of software developers and pave the way for designing machine learning based code analysis tools that can complement the existing code analysis systems for complex code queries that are either hard or expensive to represent using a formal query language. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_Mjc5NGI1OWItMmI4NC00MzAyLTgxMzAtM2IyMTA5YzhiZGE1%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%229c3e2bfe-1b0f-4d7b-a589-832878069dc6%22%7d DTSTART:20220805T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220811T120000Z UID:0dd74e279bc1988fced21af29d81d1f7-312 DTSTAMP:19700101T120018Z DESCRIPTION:Public Randomness Extraction with Ephemeral Roles and Worst-Case Corruptions URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/312/public-randomness-extraction-with-ephemeral-roles-and-worst-case-corruptions/ SUMMARY:We distill a simple information-theoretic model for randomness extraction motivated by the task of generating publicly verifiable randomness in blockchain settings and which is closely related to You-Only-Speak-Once (YOSO) protocols (CRYPTO 2021). With the goal of avoiding denial-of-service attacks, parties speak only once and in sequence by broadcasting a public value and forwarding secret values to future parties. Additionally, an unbounded adversary can corrupt any chosen subset of at most t parties. In contrast, existing YOSO protocols only handle random corruptions. As a notable example, considering worst-case corruptions allows us to reduce trust in the role assignment mechanism, which is assumed to be perfectly random in YOSO. We study the maximum corruption threshold t which allows for unconditional randomness extraction in our model: – With respect to feasibility, we give protocols for t corruptions and n = 6t + 1 or n = 5t parties depending on whether the adversary learns secret values forwarded to corrupted parties immediately once they are sent or only once the corrupted party is executed, respectively. Both settings are motivated by practical implementations of secret value forwarding. To design such protocols, we go beyond the committee-based approach that is sufficient for random corruptions in YOSO but turns out to be sub-optimal for chosen corruptions. – To complement our protocols, we show that low-error randomness extraction is impossible with corruption threshold t and n ≤ 4t parties in both settings above. This also provides a separation between chosen and random corruptions, since the latter allows for randomness extraction with close to n/2 random corruptions. Based on joint work with Jesper Buus Nielsen and Maciej Obremski. For more details: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d DTSTART:20220811T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220901T120000Z UID:b3c0d76fc954324f270f06a5625cf7e1-313 DTSTAMP:19700101T120019Z DESCRIPTION:A Moore's Law Forecast: Cloudy with a strong chance of ML URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/313/a-moores-law-forecast-cloudy-with-a-strong-chance-of-ml/ SUMMARY:Teams talk link: <a href="https://tinyurl.com/mvbjcsfu">https://tinyurl.com/mvbjcsfu</a> <br> Growing volumes of data, smarter edge devices, and new, diverse workloads are causing demand for computing to grow at phenomenal rates. At the same time, Moore's law is slowing down, stressing traditional assumptions around cheaper and faster systems every year. How do you respond to the current opportunities, exponentially increasing compute capacity at a fixed cost? Specifically, we will discuss the innovations and trends shaping the future computing landscape -- more "out-of-the-box" designs that consider the entire datacenter as a computer for custom silicon and software-defined infrastructure, broader open innovation ecosystems, and a whole lot of machine learning (ML). DTSTART:20220901T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220818T120000Z UID:3223d6c911cef022bc0dfa879ab9ac9f-314 DTSTAMP:19700101T120014Z DESCRIPTION:Reasoning about congestion control behavior URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/314/reasoning-about-congestion-control-behavior/ SUMMARY:The diversity of paths on the Internet makes it difficult for designers and operators to confidently deploy new congestion control algorithms (CCAs) without extensive real-world experiments, but such capabilities are not available to most of the networking community. And even when they are available, understanding why a CCA under-performs by trawling through massive amounts of statistical data from network connections is challenging. The history of congestion control is replete with many examples of surprising and unanticipated behaviors unseen in simulation but observed on real-world paths. In this talk I will present a simple mathematical approach to enable us to reason about congestion control behavior under such complex network phenomena. We use this approach in two ways. First, we develop CCAC (Congestion Control Anxiety Controller), a tool that uses formal verification to establish certain properties of CCAs. It is able to prove hypotheses about CCAs or generate counterexamples for invalid hypotheses. With CCAC, a designer can not only gain greater confidence prior to deployment to avoid unpleasant surprises, but can also use the counterexamples to iteratively improve their algorithm. We have modeled additive-increase/multiplicative-decrease (AIMD), Copa, and BBR with CCAC, and describe some surprising results from the exercise. Second, we tried designing a CCA that works under our network model. To our surprise, CCAC showed that all CCAs we tried suffer from starvation, an extreme form of unfairness. Further, starvation occurred under network conditions that are common on the internet. Motivated by this, we proved an impossibility result: current methods to develop delay-bounding CCAs cannot always avoid starvation. We identify a key property that makes current CCAs susceptible to starvation: when run on a path with a fixed bottleneck rate, these CCAs converge to a small delay range in equilibrium. Starvation may occur when such a CCA runs on paths that have non-congestive network delay variations due to real-world factors such as ACK aggregation and end-host scheduling. These can cause the CCA to misestimate capacity and starve. Finally I will talk about how these techniques may be used understand the performance of other heuristics used in computer systems in a mathematically precise and practically relevant way. DTSTART:20220818T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220818T120000Z UID:62c699ad477d02efd66cc9aee6cb2854-315 DTSTAMP:19700101T120018Z DESCRIPTION:Decentralized Bandits in Matching Markets URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/315/decentralized-bandits-in-matching-markets/ SUMMARY:In this talk we introduce the problem of decentralized bandits in matching markets, a model inspired by modern recommendation systems. We design algorithms for regret minimization in the two-sided matching market with one-sided bandit feedback. First, for general markets, for any ε>0, we design an algorithm that achieves a O(log1+ε(T)) regret t o the agent-optimal stable matching, with unknown time horizon T. Then, we look at a series of generalized assumptions on the market and provide algorithms that achieve the agent-optimal regret bound of O(log T). We first start with the canonical serial dictatorship setting with uniform valuations and then extend it to markets satisfying uniqueness consistency -- markets where leaving participants dont alter the original stable matching. We propose a phase-based algorithm, wherein each phase, besides deleting the globally communicated dominated arms the agents locally delete arms with which they collide often. This local deletion is pivotal in breaking deadlocks arising from rank heterogeneity of agents across arms. These phase based algorithms are ideal for deploying in a semi-online/batch setting that are common in large scale production systems. For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d DTSTART:20220818T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220822T120000Z UID:215b2dda67fa4947b6bf8f135214f075-316 DTSTAMP:19700101T120014Z DESCRIPTION:The mathematics of value-based reinforcement learning approaches URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/316/the-mathematics-of-value-based-reinforcement-learning-approaches/ SUMMARY:Reinforcement Learning (RL) studies the problem of learning in an interactive environment. In this talk, we shall discuss the main optimization problem that underlies RL and the two broad categories of approaches for solving it: value-based and policy-based. In particular, we shall focus on a popular value-based scheme called Q-learning and discuss its convergence and the main reasons behind its popularity. These discussions should provide the necessary background for the second talk. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZTk5YmQxZWItOTlmOS00YzM1LWE2ZDktNTZmNjJjNTU1OTBm%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22adc1e56f-56ee-4d24-873f-341c97ae782a%22%7d DTSTART:20220822T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220822T120000Z UID:014e5922b259997ba2400bd7a99cd164-317 DTSTAMP:19700101T120015Z DESCRIPTION:Value-based RL with function approximation and ε-greedy exploration: a differential inclusion analysis URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/317/value-based-rl-with-function-approximation-and-i%c2%b5-greedy-exploration-a-differential-inclusion-analysis/ SUMMARY:The value-based method of Q-learning with $epsilon$-greedy exploration is one of the most widely used Reinforcement Learning (RL) algorithms. While its tabular form converges to the optimal Q-function under mild conditions, the behavior of its function approximation variant is quite mysterious. Sometimes, the tactic of function approximation with greedy exploration appears to speed up learning. However, at other times, it seems to cause complex behaviors such as i.) instability, ii.) policy oscillation and chattering, iii.) multiple attractors, and iv.) worst policy convergence. Accordingly, a formal recipe to explain these phenomena has been a long-standing open problem (Sutton, 1999). In this talk, we shall provide the first pathway, based on differential inclusions, to systematically identify and explain the range of limiting phenomena that an approximate value-based RL method with greedy exploration can exhibit, thereby answering this open question. This talk is based on our recent work titled ``Approximate Q-learning and SARSA(0) under the ϵ-greedy Policy: a Differential Inclusion Analysis DTSTART:20220822T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220831T120000Z UID:a547e8cd0ede9ccf6ed7a0bffb60bbb0-318 DTSTAMP:19700101T120009Z DESCRIPTION:Achieving a practical secure non-volatile memory system with in-Memory Integrity Verification (iMIV) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/318/achieving-a-practical-secure-non-volatile-memory-system-with-in-memory-integrity-verification-imiv/ SUMMARY:Recent commercialization of Non-Volatile Memory (NVM) technology in the form of Intel Optane enables programmers to write recoverable programs. However, the data on NVM is susceptible to a plethora of data remanence attacks, which makes confidentiality and integrity protection of data essential for a secure NVM system. However, that requires computing and maintaining a large amount of security metadata (encryption counters, message authentication code (MAC), and integrity tree nodes (BMT)). Furthermore, crash consistency guarantees require the system to persist the security metadata and data atomically back to NVM, incurring high overheads. So there is a trade-off between providing complete security guarantees, the performance and recovery time of an NVM system. To ensure the confidentiality and integrity of data, a substantial quantity of security metadata is required. Of these, persisting Bonsai Merkel Tree (BMT) nodes, which are essential for fine-grain integrity verification, add substantial cost owing to the massive amount of data that must be moved off-chip to the bandwidth-constrained NVM. Thus, prior works often make a trade-off between performance and fine-grain verifiability or forego it entirely in favor of performance. The goal of this thesis is to maintain strong security and verifiability guarantees while limiting the cost of BMT updates and my thesis accomplishes this by leveraging the in-memory integrity verification. We make the fine-grain integrity verifiability realizable with a radically different approach of using in-memory computing for integrity verification. Our proposal, iMIV draws inspiration from the fact that today's commercial Optane NVM performs encryption onboard the DIMM. We argue that memory-intensive integrity verification operation should be performed near the (non-volatile) memory to avoid off-chip data movement. iMIV targets to minimize the off-chip memory transfer and mitigate the effect of bandwidth wall and also scales to larger NVM capacity in future systems with per-DIMM BMT. The experiments and analysis are carried out on a trace-driven cycle-accurate simulator VANS, which mimics the internal micro-architecture of Intel Optane memory DIMMs. The experimental results show that in comparison to the Baseline scheme with write-through caches and a strict persistency model, which also provides complete security guarantees, iMIV reduces system runtime by 1.8 times for persistent-memory aware workloads and 3.4 times for persistent-memory agnostic workloads. iMIV's recovery time on system crashes is microseconds-scale without compromising on detecting tampering and fast pin-point of the unverifiable region. iMIV limits the performance overheads associated with fine-grain integrity verifiability to less than 55 percent compared to a system that offers no security. DTSTART:20220831T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220901T120000Z UID:aa723b8cb83647c84a7c46285d19f51f-319 DTSTAMP:19700101T120009Z DESCRIPTION:Achieving a practical secure non-volatile memory system with in-Memory Integrity Verification (iMIV) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/319/achieving-a-practical-secure-non-volatile-memory-system-with-in-memory-integrity-verification-imiv/ SUMMARY:Recent commercialization of Non-Volatile Memory (NVM) technology in the form of Intel Optane enables programmers to write recoverable programs. However, the data on NVM is susceptible to a plethora of data remanence attacks, which makes confidentiality and integrity protection of data essential for a secure NVM system. However, that requires computing and maintaining a large amount of security metadata (encryption counters, message authentication code (MAC), and integrity tree nodes (BMT)). Furthermore, crash consistency guarantees require the system to persist the security metadata and data atomically back to NVM, incurring high overheads. So there is a trade-off between providing complete security guarantees, the performance and recovery time of an NVM system. To ensure the confidentiality and integrity of data, a substantial quantity of security metadata is required. Of these, persisting Bonsai Merkel Tree (BMT) nodes, which are essential for fine-grain integrity verification, add substantial cost owing to the massive amount of data that must be moved off-chip to the bandwidth-constrained NVM. Thus, prior works often make a trade-off between performance and fine-grain verifiability or forego it entirely in favor of performance. The goal of this thesis is to maintain strong security and verifiability guarantees while limiting the cost of BMT updates and my thesis accomplishes this by leveraging the in-memory integrity verification. We make the fine-grain integrity verifiability realizable with a radically different approach of using in-memory computing for integrity verification. Our proposal, iMIV draws inspiration from the fact that today's commercial Optane NVM performs encryption onboard the DIMM. We argue that memory-intensive integrity verification operation should be performed near the (non-volatile) memory to avoid off-chip data movement. iMIV targets to minimize the off-chip memory transfer and mitigate the effect of bandwidth wall and also scales to larger NVM capacity in future systems with per-DIMM BMT. The experiments and analysis are carried out on a trace-driven cycle-accurate simulator VANS, which mimics the internal micro-architecture of Intel Optane memory DIMMs. The experimental results show that in comparison to the Baseline scheme with write-through caches and a strict persistency model, which also provides complete security guarantees, iMIV reduces system runtime by 1.8 times for persistent-memory aware workloads and 3.4 times for persistent-memory agnostic workloads. iMIV's recovery time on system crashes is microseconds-scale without compromising on detecting tampering and fast pin-point of the unverifiable region. iMIV limits the performance overheads associated with fine-grain integrity verifiability to less than 55 percent compared to a system that offers no security. DTSTART:20220901T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220824T120000Z UID:60d532ac7b132f4626c58bbbfedd5fa0-322 DTSTAMP:19700101T120018Z DESCRIPTION:Local Codes for Insertion and Deletion Errors URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/322/local-codes-for-insertion-and-deletion-errors/ SUMMARY:Locally Decodable Codes (LDCs) are error-correcting codes for which individual message symbols can be quickly recovered despite errors in the codeword. LDCs for Hamming errors have been studied extensively in the past few decades, where a major goal is to understand the amount of redundancy that is necessary and sufficient to decode from large amounts of error, with small query complexity. <br> In this talk I will describe our recent results on LDCs and their variants, when the errors are in the form of insertions and deletions, rather than classical Hamming errors. Local codes against insertions and deletions are well-motivated by recent progress on DNA storage technologies. I will conclude with several intriguing open problems. (Based on joint work with many co-authors.) <br> <br> For more details please visit https://www.csa.iisc.ac.in/iisc-msr-seminar/ <br> Microsoft teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20220824T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220901T120000Z UID:3feaa69b3b3e4e8cf0476dfccaae1336-324 DTSTAMP:19700101T120018Z DESCRIPTION:Robust Secretary Algorithms for Packing Integer Programs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/324/robust-secretary-algorithms-for-packing-integer-programs/ SUMMARY:We study the problem of solving Packing Integer Programs (PIPs) in the online setting, where columns in [0,1]^d of the constraint matrix are revealed sequentially, and the goal is to pick a subset of the columns that sum to at most B in each coordinate while maximizing the objective. E.g., this problem captures the Online Knapsack problem when d=1. Excellent results are known for PIPs in the secretary model, where the columns are adversarially chosen but presented in a uniformly random order. However, these existing algorithms are susceptible to adversarial attacks: they try to DTSTART:20220901T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220908T120000Z UID:2866ec3d2e250a169b2bf4f348fd60e8-326 DTSTAMP:19700101T120014Z DESCRIPTION:A Context-Aware Neural Approach for Explainable Citation Link Prediction URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/326/a-context-aware-neural-approach-for-explainable-citation-link-prediction/ SUMMARY:Citations have become an integral part of scientific publications. They play a crucial role in supporting authors claims throughout a scientific paper. However, citing related work is a challenging and laborious task, especially for novice researchers who are not much familiar with the literature and have little or no experience in writing citation text. In this work, we study the task of Citation Link Prediction and propose a novel neural architecture called ExCite, that predicts the existence of a citation link between a pair of scientific documents within a given context. More importantly, it also generates the corresponding citation text at the same time. For this purpose, ExCite leverages diverse role-based views of the documents to learn robust document representations. The proposed model achieves state-of-the-art performance on both citation link prediction and citation text generation subtasks. We performed an extensive set of experiments to show the effectiveness of each module in the proposed neural architecture and evaluated our explanations using a wide range of state-of-the-art automatic evaluation metrics. By performing qualitative and quantitative analyses, we showed that ExCite is capable of generating high-quality citation text that is highly coherent with the citation context. DTSTART:20220908T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220912T120000Z UID:c5e9fe6e01a5449b8c9ff3d0dbc80f93-327 DTSTAMP:19700101T120019Z DESCRIPTION:Heterogeneous Systems Resilience: From Research to Industry Standards URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/327/heterogeneous-systems-resilience-from-research-to-industry-standards/ SUMMARY:Reliability is a fundamental abstraction of computing. This abstraction is increasingly challenging to achieve at high node-level component densities and for large compute infrastructures. Industry standards have played a key role in enabling such scaling, by facilitating greater heterogeneity, tighter integration of compute and memory, and paving the way for new node and system architectures. Therefore, Reliability, Availability, and Serviceability (RAS) techniques that enhance resilience and intercept major industry standards are beneficial to the overall ecosystem using these standards. We first explain why RAS is important for large-scale systems and outline some key RAS best practices for servers. We then present insights from studying reliability field data from production systems in data centers and an overview of tools and techniques developed to enhance resiliency and reliability. Finally, we show how the research influenced the RAS architecture and capabilities of two recent industry standards and their potential resilience benefits at scale. MS Teams link: https://tinyurl.com/yumys6wt DTSTART:20220912T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220909T120000Z UID:a6d77130d8c6d554dc6012b25a1fc5f2-329 DTSTAMP:19700101T120011Z DESCRIPTION:On High Dimensional Expanders and Hardness of Approximation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/329/on-high-dimensional-expanders-and-hardness-of-approximation/ SUMMARY:Expander graphs are a classical tool in theoretical computer science, with applications ranging from network stability all the way to the PCP Theorem. In this talk, we overview recent developments in the nascent theory of high dimensional expansion, a generalization of expanders to hypergraphs and posets, and discuss a recent line of work towards their application to hardness of approximation. We focus in particular on two interconnected narratives. First, we will cover recently popularized spectral notions of high dimensional expansion, their relation to powerful tools such as hypercontractivity, and discuss how such connections could open a new path towards proving Khots Unique Games Conjecture. Second, we will introduce lesser-known topological notions of high dimensional expansion and discuss how they lead to explicit lower bounds for fundamental problems within the Sum-of-Squares semidefinite programming hierarchy, the most powerful known algorithmic paradigm for combinatorial approximation problems such as unique games. Based on a series of joint works with Mitali Bafna, Jason Gaitonde, Tali Kaufman, Ting-Chun Lin, Shachar Lovett, and Ruizhe Zhang. For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20220909T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220915T120000Z UID:26118c86500b668c19b06e38a059fab7-330 DTSTAMP:19700101T120015Z DESCRIPTION:Building a GPU-powered Spatial Query Engine URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/330/building-a-gpu-powered-spatial-query-engine/ SUMMARY:Given the massive growth in the volume of spatial data, there is a great need for systems that can efficiently evaluate spatial queries over large data sets. These queries are notoriously expensive using traditional database solutions. While faster response times can be attained through powerful clusters or servers with large main memory, these options, due to cost and complexity, are out of reach to many data scientists and analysts making up the long tail. Graphics Processing Units (GPUs), which are now widely available even in commodity laptops, provide a cost-effective alternative to support high-performance computing, opening up new opportunities to the efficient evaluation of spatial queries. While GPU-based approaches proposed in the literature have shown great improvements in performance, they are tied to specific GPU hardware and only handle specific queries over fixed geometry types. As a first step towards making GPU spatial query processing mainstream, we propose a new model that uniformly represents spatial data as geometric objects and define an algebra consisting of GPU-friendly composable operators that operate over these objects. This is then used to develop SPADE, a spatial query engine that uses the computer graphics pipeline to realize this algebra and data model to achieve both efficiency and portability across different GPU hardware. In this talk, I will briefly go over the algebra, and discuss the implementation and efficiency of the SPADE system. DTSTART:20220915T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220912T120000Z UID:1c15b4780202f56e948d5ffcef24b8b1-331 DTSTAMP:19700101T120014Z DESCRIPTION:Model-based Safe Deep Reinforcement Learning and Empirical Analysis of Safety via Attribution URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/331/model-based-safe-deep-reinforcement-learning-and-empirical-analysis-of-safety-via-attribution/ SUMMARY:During initial iterations of training in most Reinforcement Learning (RL) algorithms, agents perform a significant number of random exploratory steps, which in the real-world limit the practicality of these algorithms as this can lead to potentially dangerous behavior. Hence safe exploration is a critical issue in applying RL algorithms in the real world. This problem is well studied in the literature under the Constrained Markov Decision Process (CMDP) Framework, where in addition to single-stage rewards, state transitions receive single-stage costs as well. The prescribed cost functions are responsible for mapping undesirable behavior at any given time-step to a scalar value. Then we aim to find a feasible policy that maximizes reward returns and keeps cost returns below a prescribed threshold during training as well as deployment. We propose a novel On-policy Model-based Safe Deep RL algorithm in which we learn the transition dynamics of the environment in an online manner as well as find a feasible optimal policy using Lagrangian Relaxation-based Proximal Policy Optimization. This combination of transition dynamics learning, and a safety-promoting RL algorithm leads to ~3-4 times less environment interactions and less cumulative hazard violations compared to the model-free approach. We use an ensemble of neural networks with different initializations to tackle epistemic and aleatoric uncertainty issues faced during environment model learning. We present our results on a challenging Safe Reinforcement Learning benchmark - the Open AI Safety Gym. In addition to this, we perform an attribution analysis of actions taken by the Deep Neural Network-based policy at each time step. This analysis helps us to : Identify the feature in state representation which is significantly responsible for the current action. Empirically provide the evidence of the safety-aware agents ability to deal with hazards in the environment provided that hazard information is present in the state representation. In order to perform the above analysis, we assume state representation has meaningful information about hazards and goals. Then we calculate an attribution vector of the same dimension as state using a well-known attribution technique known as Integrated Gradients. The resultant attribution vector provides the importance of each state feature for the current action. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZmVlMjcwNTMtNjJhYS00MGE2LWI0NDktNGEzZmU1OGFhOTZj%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2271844033-661c-432d-9a6f-418de5b8c819%22%7d DTSTART:20220912T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220916T120000Z UID:55602bee0c494644ce5ab2321532078d-332 DTSTAMP:19700101T120011Z DESCRIPTION:Correlated Rounding for Correlation Clustering URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/332/correlated-rounding-for-correlation-clustering/ SUMMARY:Given a complete graph G = (V, E) where each edge is labeled + or -, the correlation clustering problem asks to partition V into clusters to minimize the number of +edges between different clusters plus the number of -edges within the same cluster. The approximability of correlation clustering has been actively investigated [BBC04, CGW05, ACN08], culminating in a 2.06-approximation algorithm [CMSY15], based on rounding the standard LP relaxation. Since the integrality gap for this formulation is 2, it has remained a major open question to determine if the approximation factor of 2 can be reached, or even breached. In this talk, we show how to achieve a factor of 2+eps based on O(1/eps^2) rounds of the Sherali-Adams hierarchy. To round this relaxation, we adapt the correlated rounding originally developed for CSPs [BRS11, GS11, RT12]. To go below this approximation ratio, we go beyond the traditional triangle-based analysis by employing a global charging scheme that amortizes the total cost of the rounding across different triangles. Joint work with Vincent Cohen-Addad and Euiwoong Lee. For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20220916T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220919T120000Z UID:71d8c1b0b67ec5056b690269a672350e-333 DTSTAMP:19700101T120014Z DESCRIPTION:Building Storage Systems for New Applications and New Hardware URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/333/building-storage-systems-for-new-applications-and-new-hardware/ SUMMARY:The modern storage landscape is changing at an exciting rate. New technologies, such as CXL memory devices, are being introduced. At the same time, new applications such as blockchain are emerging with new requirements from the storage subsystem. New regulations, such as the General Data Protection Regulation (GDPR), place new constraints on how data may be read and written. As a result, designing storage systems that satisfy these constraints is interesting and challenging. In this talk, I will describe the lessons we learnt from tackling this challenge in various forms: my group has built file systems and concurrent data structures for persistent memory, storage solutions for blockchains and machine learning, and analyzed how the GDPR regulation affects storage systems. DTSTART:20220919T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220929T120000Z UID:12b27524151bf6d6e787ee741a563cdb-334 DTSTAMP:19700101T120014Z DESCRIPTION:An Evaluation of Basic Protection Mechanisms in Financial Apps on Mobile Devices URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/334/an-evaluation-of-basic-protection-mechanisms-in-financial-apps-on-mobile-devices/ SUMMARY:Mobile devices have become an integral part of the payment ecosystem. Payments are facilitated by financial applications (like Mobile Banking, UPI Apps, etc.), which have in turn soared in popularity. With the increasing dependence on the financial app ecosystem and the sensitive nature of the data handled by financial apps (including the bank/card details of the payees and the payers), we set out to study fundamental question: do the app developers of financial apps put various self-defense checks to make their apps more secure? If yes, how trivial is it for the attackers to bypass such checks? This thesis concerns the robustness of security checks in financial mobile applications. The best practices recommended by the Open Web Application Security Project (OWASP) for developing such apps, demand that developers include several checks in these apps, such as detection of running on a rooted device, certificate checks, and so on. Ideally, these checks must be introduced in a sophisticated way and must not be locatable through trivial static analysis, so that attackers cannot bypass them trivially. In this work, we conduct a large-scale study focused on financial apps on the Android platform and determine the robustness of these checks. Our study shows that a significant fraction of the financial apps dont have the various self-defense checks recommended by the OWASP. Then we showed that among the apps with at least one security check, > 50% of such apps at least one check could be trivially bypassed. Some of such financial apps have installation counts exceeding 100 million from Google Play. This entire process of detecting the self-defense check and bypassing it is automated. We believe that the results of our study can guide developers of these apps in inserting security checks in a more robust fashion. DTSTART:20220929T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221010T120000Z UID:89864477cf4b8031e1d2dbf3352bcc60-335 DTSTAMP:19700101T120015Z DESCRIPTION:Topological Data Analysis, Basics, Computation and Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/335/topological-data-analysis-basics-computation-and-applications/ SUMMARY:In this talk, we will discuss the basic theory of Topological data analysis (TDA), in particular, Persistent Homology (PH). Then we will look into its computational aspects including the challenges and the recent advancements. We will also discuss our recent work (SoCG 22) with Marc Gliss to speed-up the computation of flag filtration using DTSTART:20221010T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220923T120000Z UID:ea22e7631e4c468d6cdddea4a7871e47-336 DTSTAMP:19700101T120011Z DESCRIPTION:Approximability of p q Matrix Norms:Some NP-hardness Results and An Approximation Algorithm URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/336/approximability-of-p-q-matrix-normssome-np-hardness-results-and-an-approximation-algorithm/ SUMMARY:A real n x m matrix A represents a linear map from $R^m$ to $R^n$. If the domain and the range of the map is endowed with $ell_p$-norm and -norm respectively, then we can measure the &quot;stretch&quot; of the map A. Thus matrix p-&gt;q-norm of a matrix A is defined to be $sup_{xin R^n, ||x||_ple 1} || Ax ||_q$. This quantity generalizes the well known spectral norm which is very useful in many areas of theoretical computer science (and that of mathematics). This can also be seen as a special case of optimizing a polynomial over a convex body. When $p&lt;q$, we call this problem hypercontractive matrix norm estimation, and the case of $p ge q$ is called reverse-hypercontractive matrix norm estimation. The placement of 2 with respect to the interval $p$ to $q$ turns out to be of importance in approximability or inapproximability of the problem. The case of $infty to 1$-norm was studied by Grothendieck in his famous Resume (1956). The integrality gap of an SDP relaxation for this problem is a constant, called Grothendiecks constant. The combinatorial quantity cut-norm of a graph is related to the $infty to 1$-norm of the adjacency matrix of the graph, as observed by Alon and Naor (2004). Therefore, this problem has relevance in combinatorial optimization problems . NP-hardness of the $inftyto 1$ problem was established by Briët, Regev and Saket (2015) within the constant $pi/2$ for the special case of positive semidefinite matrix. Constant factor approximation algorithms for the case $2in [q,p]$ were developed by Nesterov (1998) and Steinberg (2005). For the case $2 not in [q,p]$, quasi-polynomial hardness was shown by Bhaskara and Vijayaraghavan (2015) assuming NP does not have quasi-polynomial time algorithms. The problem of the $2to 4$ norm is relevant in several problems in quantum information theory (as shown in work of Harrow and Montanaro (2013)). Barak et.al. (2012) showed that the $2to q$-norm estimation problem (for even $qge 4$) is intimately related to the problem of Small Set Expansion. They also proved a hardness of approximation assuming Exponential Time Hypothesis. We proved the first NP hardness of the hypercontractive norm in case $2notin [p,q]$. We show quasi-polynomial hardness of approximation assuming NP does not have quasi-polynomial time randomized algorithms. For the reverse-hypercontractive case of $2in [q,p]$, we proved NP hardness of approximation within a constant ($1/(gamma_q gamma_{p*})$) factor. In an associated work we gave an algorithm with approximation factor constant away from the above -- that is the dependency on $p$ and $q$ on the hardness factor is roughly correct. We also recovered the hardness result of Bhaskara and Vijayaraghavan (2015) for the reverse-hypercontractive $2 not in [q,p]$ case, albeit under slightly stronger hardness assumption. All of our hardness results follow a general framework. Based on joint work with Vijay Bhattiprolu, Venkatesan Guruswami, Euiwoong Lee, and Madhur Tulsiani. Appeared in SODA 2019. For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d DTSTART:20220923T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220926T120000Z UID:faa25332dfd6e9d8f403413e63d96287-337 DTSTAMP:19700101T120010Z DESCRIPTION:Fast Algorithms for Max Cut on Geometric Intersection Graphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/337/fast-algorithms-for-max-cut-on-geometric-intersection-graphs/ SUMMARY:In the max cut problem, given a graph, the goal is to partition the vertex set into two disjoint sets such that the number of edges having their endpoints in different sets is maximized. Max cut is an NP-hard problem. The seminal work by Goemans and Williamson gave an approximation algorithm for the max cut problem having an approximation ratio of 0.878. In this work, we design fast algorithms for max cut on geometric intersection graphs. In a geometric intersection graph, given a collection of n geometric objects as the input, each object corresponds to a vertex and there is an edge between two vertices if and only if the corresponding objects intersect. Since we are dealing with the geometric intersection graphs, which have more structure than general graphs, the following questions are of interest: 1. Are there special cases of geometric intersection graphs for which max cut can be solved exactly in polynomial time? 2. It can be shown that the random cut gives a 0.5 approximation for the max cut. Is it possible to design linear or near-linear time algorithms (in terms of n) and beat the 0.5 approximation barrier? The edge-set of the graph is not explicitly given as input; therefore, designing linear time algorithms is of interest. 3. Can an approximation factor better than 0.878 be obtained for the geometric intersection graphs? We obtain the following results: An exact and fast algorithm for laminar geometric intersection graphs. Our algorithm uses a greedy strategy. A fast algorithm is obtained by combining the properties of laminar objects with range searching data structures. An O(n log n) time algorithm with an approximation factor of 2/3 for unit interval intersection graphs. We decompose the unit intervals into several cliques, and based on the number of edges between adjacent cliques, we choose an appropriate partitioning strategy. An O(n log n) time algorithm with an approximation factor of 7/13 for unit square intersection graphs. We use the largest clique in the graph to beat the 0.5 approximation barrier. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_NDY1ZjkzYzEtMzEwMC00OGU5LTk2NDQtMDQ3ZDAzYmU0M2I2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d71644d4-9e9d-48e9-ab41-3a147899b402%22%7d DTSTART:20220926T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220927T120000Z UID:840495d0dbe5d689996eb95cf40d4304-338 DTSTAMP:19700101T120011Z DESCRIPTION:Top-k Spatial Aware Ads URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/338/top-k-spatial-aware-ads/ SUMMARY:Consider an app on a smartphone which displays local business ads. When a user opens the app, then k local business ads need to be displayed (where k would typically be 3 or 5) such that the profit made by the app is maximized. The pricing model needs to consider that (a) each business is willing to bid a different price, and (b) farther the distance of the user on whose smartphone the ad is displayed, the lesser is the price paid by to the app. Motivated by such applications, in this work, we design fast algorithms to retrieve top-k objects using the provided spatial and non-spatial attributes. We refer them as Top-k Spatial Aware Ads Queries (SAA). In Top-k-Saa, the query is user location, and we return top-k objects that have the best score. The scoring function is based on the distance between the object and query point (spatial attribute) and non-spatial attributes. We propose algorithms that efficiently preprocess the data using appropriate data structures and aid in fast query processing. A simple O(n log k) algorithm returns the top-k ads based on the scoring function value. We obtain the following results. Our first algorithm uses O(n log n) space and answers the top-k SAA query in O(k log2n) time. The fast query time is obtained by leveraging the properties of additively weighted Voronoi diagram, along with other supporting data structures. Our second algorithm improves upon the first algorithm by improving the query time to O(k log n), while using the same space. This is achieved via an interesting combination of randomization with a top-2 structure. The proposed algorithms have good worst-case theoretical bounds. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZmUxNWM1ODgtYWVmNi00MWRmLTk0MDMtZDA0Yzk2OTRjNzBk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22680e1edf-a4ab-4b4e-9ddb-38c0f9d5ad80%22%7d DTSTART:20220927T120000Z END:VEVENT BEGIN:VEVENT DTEND:20220930T120000Z UID:f61c66b250bbe5b73023d4b0ca8b769b-339 DTSTAMP:19700101T120016Z DESCRIPTION:Negative-Weight Single-Source Shortest Paths in Near-linear Time URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/339/negative-weight-single-source-shortest-paths-in-near-linear-time/ SUMMARY:We present a randomized algorithm that computes single-source shortest paths (SSSP) in O(m log^8(n) log W) time when edge weights are integral and can be negative and are >=-W. This essentially resolves the classic negative-weight SSSP problem. In contrast to all recent developments that rely on sophisticated continuous optimization methods and dynamic algorithms, our algorithm is simple: it requires only a simple graph decomposition and elementary combinatorial tools. Organizers note: The talk is based on a paper that is the co-winner of the best paper award at FOCS 22. For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20220930T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221007T120000Z UID:c9d2e34d2c343acb0d21e3745a85c8df-340 DTSTAMP:19700101T120016Z DESCRIPTION:Improved sublinear algorithms for testing permutation freeness URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/340/improved-sublinear-algorithms-for-testing-permutation-freeness/ SUMMARY:Given a permutation pi of length k, an array A is pi-free if there are no k array values that, when considered from left to right, have the same relative order as that of the permutation. For example, the array A has is (2,1)-free if there are no two indices i < j such that A[i] > A[j] and the array is (1,3,2)-free if there are no three indices i < j < k such that A[j] > A[k] > A[i]. In particular, the set of (2,1)-free arrays are simply the set of all sorted arrays. The problem of testing pi-freeness is to distinguish arrays that are pi-free from arrays that need to be modified in at least a constant fraction of their values to be pi-free. This problem was first studied systematically by Newman, Rabinovich, Rajendra Prasad and Sohler (Random Structures and Algorithms; 2019), where they proved that for all permutations pi of length at most 3, one can test for pi-freeness in polylog n many queries, where n is the array length. For permutations of length k > 3, they described a simple testing algorithm that makes O(n^{1-1/k}) queries. Most of the followup work has focused on improving the query complexity of testing pi-freeness for monotone pi. In this talk, I will present a recent algorithm with query complexity O(n^{o(1)}) that tests pi-freeness for arbitrary permutations of constant length, which significantly improves the state of the art. I will also give an overview of the analysis that involves several combinatorial ideas. <br> Joint work with Ilan Newman <br> Organizers Note: This paper had won the best paper award at ICALP 2022. <br> <br> For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ <br> <br> Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20221007T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221015T120000Z UID:c1e6c19182fb1c4b2446f9b3c2404842-341 DTSTAMP:19700101T120014Z DESCRIPTION:Reinforcement Learning via Stochastic Approximation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/341/reinforcement-learning-via-stochastic-approximation/ SUMMARY:Reinforcement Learning (RL) is one of the most popular branches of AI/ML. Moreover, many though not all of the currently popular RL algorithms have a solid mathematical justification. At present, a wide variety of proof methods are used to study RL algorithms. Therefore, there is scope for some unification of these diverse approaches. In this talk, I will show how a large fraction of current RL algorithms can be viewed as implementations of the Stochastic Approximation (SA) algorithm, in various forms. I will also show that, by using martingale methods, the convergence of these SA variants can be established more simply and with fewer assumptions than is possible using the ODE approach, which is another popular method of analyzing SA algorithms. Some problems for future research will also be indicated. DTSTART:20221015T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221014T120000Z UID:0ec92cc3bbaec65fb1777dd7f5ccf596-342 DTSTAMP:19700101T120011Z DESCRIPTION:Determinant Maximization via Matroid Intersection Algorithms URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/342/determinant-maximization-via-matroid-intersection-algorithms/ SUMMARY:Determinant maximization problem gives a general framework that models problems arising in as diverse fields as statistics, convex geometry, fair allocations, combinatorics, spectral graph theory, network design, and random processes. In an instance of a determinant maximization problem, we are given a collection of vectors $U = {v_1, ldots, v_n}$ in $d$ dimensions, and a goal is to pick a subset S ⊆ U of given vectors to maximize the determinant of the matrix $sum_{i in S} v_i v_i^T$. Often, the set $S$ of picked vectors must satisfy additional combinatorial constraints such as cardinality constraint ($|S| leq k$) or matroid constraint ($S$ is a basis of a matroid defined on the vectors). In this talk, we give a polynomial-time deterministic algorithm that returns an $r^{O(r)}$-approximation for any matroid of rank $r leq d$. Our algorithm builds on combinatorial algorithms for matroid intersection, which iteratively improves any solution by finding an alternating negative cycle in the exchange graph defined by the matroids. While the determinant function is not linear, we show that taking appropriate linear approximations at each iteration suffice to give the improved approximation algorithm. Organizers Note: Joint work with Adam Brown, Madhusudhan Pittu, Mohit Singh, and Prasad Tetali. For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20221014T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221014T120000Z UID:354f8fd503df8fef2ef3229b612b9ce3-343 DTSTAMP:19700101T120011Z DESCRIPTION:Fully-Secure MPC with Minimal Trust URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/343/fully-secure-mpc-with-minimal-trust/ SUMMARY:The task of achieving full security (with guaranteed output delivery) in secure multiparty computation (MPC) is a long-studied problem. Known impossibility results (Cleve, STOC 86) rule out general solutions in the dishonest majority setting. In this work, we consider solutions that use an external trusted party (TP) to bypass the impossibility results, and study the minimal requirements needed from this trusted party. In particular, we restrict ourselves to the extreme setting where the size of the TP is independent of the size of the functionality to be computed (called “small” TP) and this TP is invoked only once during the protocol execution. We present several positive and negative results for fully-secure MPC in this setting. -- For a natural class of protocols, specifically, those with a universal output decoder, we show that the size of the TP must necessarily be exponential in the number of parties. This result holds irrespective of the computational assumptions used in the protocol. We additionally rule out the possibility of achieving information-theoretic full security (without the restriction of using a universal output decoder) using a “small” TP in the plain model (i.e., without any setup). -- In order to get around the above negative result, we consider protocols without a universal output decoder. The main positive result in our work is a construction of such a fully-secure MPC protocol assuming the existence of a succinct Functional Encryption scheme. -- Finally, we explore the possibility of achieving full-security with a semi-honest TP that could collude with other malicious parties (which form a dishonest majority). In this setting, we show that even fairness is impossible to achieve regardless of the “small TP” requirement. DTSTART:20221014T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221019T120000Z UID:d2e8ffb9389f7897162bb3ee904203a2-344 DTSTAMP:19700101T120011Z DESCRIPTION:Multi-Armed Bandits – On Range Searching and On Slowly-varying Non-stationarity. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/344/multi-armed-bandits-ae-on-range-searching-and-on-slowly-varying-non-stationarity/ SUMMARY:Multi-Armed Bandits (MAB) is a popular framework for modelling sequential decision-making problems under uncertainty. This thesis is a compilation of two independent works on MABs. 1. Optimal Algorithms for Range Searching over Multi-Armed Bandits We consider a multi-armed bandit (MAB) version of the range-searching problem. In its basic form, range searching considers as input a set of points (on the real line) and a collection of (real) intervals. Here, with each specified point, we have an associated weight, and the problem objective is to find a maximum-weight point within every given interval. The current work addresses range searching with stochastic weights: each point corresponds to an arm (that admits sample access) and the points weight is the (unknown) mean of the underlying distribution. In this MAB setup, we develop sample-efficient algorithms that find, with high probability, near-maximum-weight points (arms) within the given intervals, i.e., we obtain PAC (probably approximately correct) guarantees. We also provide an algorithm for a generalization wherein the weight of each point is a multi-dimensional vector. The sample complexities of our algorithms depend, in particular, on the size of the optimal hitting set of the given intervals. Finally, we establish lower bounds proving that the obtained sample complexities are essentially tight. Our results highlight the significance of geometric constructs -- specifically, hitting sets -- in our MAB setting. 2. On Slowly-varying Non-stationary Bandits We consider minimisation of dynamic regret in non-stationary bandits with a slowly varying property. Namely, we assume that arms rewards are stochastic and independent over time, but that the absolute difference between the expected rewards of any arm at any two consecutive time-steps is at most a drift limit δ>0. For this setting that has not received enough attention in the past, we give a new algorithm which extends naturally the well-known Successive Elimination algorithm to the non-stationary bandit setting. We establish the first instance-dependent regret upper bound for slowly varying non-stationary bandits. The analysis in turn relies on a novel characterization of the instance as a detectable gap profile that depends on the expected arm reward differences. We also provide the first minimax regret lower bound for this problem, enabling us to show that our algorithm is essentially minimax optimal. Also, this lower bound we obtain matches that of the more general total variation-budgeted bandits problem, establishing that the seemingly easier former problem is at least as hard as the more general latter problem in the minimax sense. We complement our theoretical results with experimental illustrations. Microsoft Teams Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_M2IzZTA4ZmItZGMwZi00MzM4LWFhOGEtMGY1NjAyNWVhMjUw%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22146013fa-e7e4-4c88-8bee-e6f407d8db1e%22%7d DTSTART:20221019T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221020T120000Z UID:52098c48620106937d911f868db94be1-345 DTSTAMP:19700101T120017Z DESCRIPTION:Linear-map Vector Commitments and their Practical Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/345/linear-map-vector-commitments-and-their-practical-applications/ SUMMARY:Vector commitments (VC) are a cryptographic primitive that allow one to commit to a vector and then “open” some of its positions or a function of them efficiently. Vector commitments are increasingly recognized as a central tool to scale highly decentralized networks of large size and whose content is dynamic. In this work, we examine the demands on the properties that an ideal vector commitment should satisfy in the light of the emerging plethora of practical applications and propose new constructions that improve the state-of-the-art in several dimensions and offer new tradeoffs. We also propose a unifying framework for functional openings that captures several constructions and shows how to generically achieve some properties from more basic ones. On the practical side, we focus on building schemes that over-perform in efficiency prior schemes and do not require new trusted setup (we can reuse existing ceremonies for pairing-based “powers of tau” run by real-world systems such as ZCash or Filecoin). DTSTART:20221020T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221021T120000Z UID:0793af7546173bdb9beff941f2d61158-346 DTSTAMP:19700101T120015Z DESCRIPTION:Improved Algorithms for Variants of Bin Packing and Knapsack. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/346/improved-algorithms-for-variants-of-bin-packing-and-knapsack/ SUMMARY:We study variants of two classical optimization problems: Bin Packing and Knapsack. Both bin packing and knapsack fall under the regime of DTSTART:20221021T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221021T120000Z UID:ccce190ab0f6e57df59e945731cf44aa-347 DTSTAMP:19700101T120015Z DESCRIPTION:Improved Algorithms for Variants of Bin Packing and Knapsack. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/347/improved-algorithms-for-variants-of-bin-packing-and-knapsack/ SUMMARY:We study variants of two classical optimization problems: Bin Packing and Knapsack. Both bin packing and knapsack fall under the regime of DTSTART:20221021T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221027T120000Z UID:e3bf10bf9f390e3e5f202030f948b230-348 DTSTAMP:19700101T120011Z DESCRIPTION:Dynamic Data Race Prediction: Fundamentals and Advances URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/348/dynamic-data-race-prediction-fundamentals-and-advances/ SUMMARY:Concurrent programs are notoriously hard to write correctly, as scheduling nondeterminism introduces subtle errors that are both hard to detect and to reproduce. Data races are arguably the most insidious amongst concurrency bugs and extensive research efforts have been dedicated to effectively detect them. A data race occurs when memory-conflicting actions are executed concurrently. Consequently, considerable effort has been made towards developing efficient techniques for race detection. The preferred approach to detect data races is through dynamic analysis, where one observes an execution of a concurrent program and checks for the presence of data races in the execution observed. Traditional dynamic race detectors rely on Lamport's happens-before (HB) partial order, which can be conservative and are often unable to discover simple data races, even after executing the program several times. Dynamic data race prediction aims to expose data races, that can be otherwise missed by traditional dynamic race detectors (such as those based on HB), by inferring data races in alternate executions of the underlying program, without re-executing it. In this talk, I will talk about the fundamentals of and recent algorithmic advances in data race prediction. DTSTART:20221027T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221028T120000Z UID:577380d95256cd6851ea0740074226f8-349 DTSTAMP:19700101T120011Z DESCRIPTION:Set membership with two classical and quantum probes URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/349/set-membership-with-two-classical-and-quantum-probes/ SUMMARY:We consider the following data structure problem. Given an n-element subset S of a universe of size m, represent S in memory as a bit string x(S) so that membership queries of the form DTSTART:20221028T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221111T120000Z UID:9a90b1876a58ab742a6dcd399f84e064-351 DTSTAMP:19700101T120011Z DESCRIPTION:Rethinking the IID Assumption for Data-efficient and Robust NLP URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/351/rethinking-the-iid-assumption-for-data-efficient-and-robust-nlp/ SUMMARY:The standard practice for training Machine Learning models is to assume access to independent and identically distributed (IID) labelled data, to optimise for the average loss on a training set and to measure generalisation on a held-out in-distribution test set. While this assumption is grounded in theory, it does not often hold in practice. In this talk I will focus on two issues with the setup that affect NLP systems: 1) Obtaining sufficient high quality labelled training data can be expensive, making it difficult to train models that generalise well to in-distribution test sets; 2) Even when large enough labelled training sets are available, they may come with unwanted correlations between labels and task-independent features. These spurious correlations are a consequence of unavoidable biases in the dataset collection processes, and can provide shortcuts to models that allow them to generalise well to in-distribution test sets that also have those spurious correlations, without actually learning the intended tasks. To illustrate the first problem, I will first present Qasper, a complex document-level Question Answering task over research papers and an associated dataset collection process that requires expert annotators. To address the problem, I will present a data-efficient training method that leverages data from other tasks where it is easier to obtain labelled data. In contrast with recent work that trained massive multitask models (e.g. T0, FLAN) on tens of millions of instances from all available datasets without any knowledge of the target tasks, our method selects small subsets of multi-task training instances that are relevant to the target tasks, using only unlabelled target task instances. Our method is algorithmically simple, yet quite effective and data-efficient—on Qasper and ten other datasets, the target-task specific models outperform the T0 model of the same size by up to 30% without accessing target task labels (zero-shot), and by up to 23% while accessing a few target-task labels (few-shot), all while using about 2% of the multitask data used to train T0 models. To address the second problem, as an alternative to Empirical Risk Minimisation (ERM) which optimises for average training loss under the IID assumption, I will present a novel optimisation technique based on Group Distributionally Robust Optimisation (G-DRO) which optimises for worst group performance over a known set of pre-defined groups in the training data. Since spurious correlations are often unknown, directly applying G-DRO to this problem is not feasible. Our method, AGRO, simultaneously identifies error-prone groups in the training data and optimises for the model’s performance on them. We show that on several NLP and Vision tasks, AGRO based models outperform models trained using ERM on known error-prone subsets of in-distribution test data by up to 8%, and also on out-of-distributions test sets by up to 10% without using any knowledge of the distribution shifts. DTSTART:20221111T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221028T120000Z UID:5a62feef7d3130cf9ae4c44fd44f37ee-352 DTSTAMP:19700101T120011Z DESCRIPTION:Dynamic Data Race Prediction: Fundamentals and Advances URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/352/dynamic-data-race-prediction-fundamentals-and-advances/ SUMMARY:Please note that this talk is rescheduled to 11:30am on Friday 28th October in Room 252 CSA Department. Apologies for the inconvenience caused. ----------------------------------------------------------------------------------------------------------------------- Concurrent programs are notoriously hard to write correctly, as scheduling nondeterminism introduces subtle errors that are both hard to detect and to reproduce. Data races are arguably the most insidious amongst concurrency bugs and extensive research efforts have been dedicated to effectively detect them. A data race occurs when memory-conflicting actions are executed concurrently. Consequently, considerable effort has been made towards developing efficient techniques for race detection. The preferred approach to detect data races is through dynamic analysis, where one observes an execution of a concurrent program and checks for the presence of data races in the execution observed. Traditional dynamic race detectors rely on Lamport's happens-before (HB) partial order, which can be conservative and are often unable to discover simple data races, even after executing the program several times. Dynamic data race prediction aims to expose data races, that can be otherwise missed by traditional dynamic race detectors (such as those based on HB), by inferring data races in alternate executions of the underlying program, without re-executing it. In this talk, I will talk about the fundamentals of and recent algorithmic advances in data race prediction. DTSTART:20221028T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221103T120000Z UID:58b37ffd8c713326a927211de63eb9d6-353 DTSTAMP:19700101T120017Z DESCRIPTION:Asymptotically Free Broadcast in Constant Expected Time via Packed VSS URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/353/asymptotically-free-broadcast-in-constant-expected-time-via-packed-vss/ SUMMARY:Broadcast is an essential primitive for secure computation. We focus in this paper on optimal resilience (i.e., when the number of corrupted parties t is less than a third of the computing parties n), and with no setup or cryptographic assumptions. While broadcast with worst case t rounds is impossible, it has been shown [Feldman and Micali STOC88, Katz and Koo CRYPTO06] how to construct protocols with expected constant number of rounds in the private channel model. However, those constructions have large communication complexity, specifically O(n2L + n6log n) expected number of bits transmitted for broadcasting a message of length L. This leads to a significant communication blowup in secure computation protocols in this setting. In this paper, we substantially improve the communication complexity of broadcast in constant expected time. Specifically, the expected communication complexity of our protocol is O(nL + n4log n). For messages of length L = Ω(n3log n), our broadcast has no asymptotic overhead (up to expectation), as each party has to send or receive O(n3log n) bits. We also consider parallel broadcast, where n parties wish to broadcast L bit messages in parallel. Our protocol has no asymptotic overhead for L = Ω(n2log n), which is a common communication pattern in perfectly secure MPC protocols. For instance, it is common that all parties share their inputs simultaneously at the same round, and verifiable secret sharing protocols require the dealer to broadcast a total of O(n2log n) bits. As an independent interest, our broadcast is achieved by a packed verifiable secret sharing, a new notion that we introduce. We show a protocol that verifies O(n) secrets simultaneously with the same cost of verifying just a single secret. This improves by a factor of n the state-of-the-art. DTSTART:20221103T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221110T120000Z UID:967f3509d52d205f0c55ddb434248821-354 DTSTAMP:19700101T120021Z DESCRIPTION:Cheeger Inequalities for Vertex Expansion and Reweighted Eigenvalues URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/354/cheeger-inequalities-for-vertex-expansion-and-reweighted-eigenvalues/ SUMMARY:The classical Cheegers inequality relates the edge conductance ψ of a graph and the second smallest eigenvalue λ2 of the Laplacian matrix. Recently, Olesker-Taylor and Zanetti discovered a Cheeger-type inequality ψ2 / log |V| ≲ λ2* ≲ ψ connecting the vertex expansion ψ of a graph G=(V,E) and the maximum reweighted second smallest eigenvalue λ2* of the Laplacian matrix. In this work, we first improve their result to ψ2 / log d ≲ λ2* ≲ ψ where d is the maximum degree in G, which is optimal up to a constant factor. Also, the improved result holds for weighted vertex expansion, answering an open question by Olesker-Taylor and Zanetti. Building on this connection, we then develop a new spectral theory for vertex expansion. We discover that several interesting generalizations of Cheeger inequalities relating edge conductances and eigenvalues have a close analog in relating vertex expansions and reweighted eigenvalues. These include an analog of Trevisans result on bipartiteness, an analog of higher order Cheegers inequality, and an analog of improved Cheegers inequality. Finally, inspired by this connection, we present negative evidence to the 0/1-polytope edge expansion conjecture by Mihail and Vazirani. We construct 0/1-polytopes whose graphs have very poor vertex expansion. This implies that the fastest mixing time to the uniform distribution on the vertices of these 0/1-polytopes is almost linear in the graph size. This does not provide a counterexample to the conjecture, but this is in contrast with known positive results which proved poly-logarithmic mixing time to the uniform distribution on the vertices of subclasses of 0/1-polytopes. Speaker website: https://cs.uwaterloo.ca/~lapchi/ For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20221110T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221118T120000Z UID:7e58fd5f074aa3b84b7dc64eb19fd8ab-355 DTSTAMP:19700101T120010Z DESCRIPTION:Efficient zero-knowledge proofs based on vector-oblivious linear evaluation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/355/efficient-zero-knowledge-proofs-based-on-vector-oblivious-linear-evaluation/ SUMMARY:Zero-knowledge (ZK) proofs with an optimal memory footprint have attracted a lot of attention because such protocols can easily prove very large computations with a small memory requirement. In this talk, the speaker will talk about some recent progress on concretely efficient ZK protocols based on VOLE and their applications in this setting. These protocols are very cheap computationally and can prove large statements like ResNet inference or large RAM-based computation with ease; on the other hand, it is designated-verifier, and the proof size is often linear to the circuit size. Finally, the speaker will talk about more recent advances that lead to sublinear communication VOLE zero-knowledge proof protocols. DTSTART:20221118T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221124T120000Z UID:773ae399b7d59a6d8bde02c610f42dc8-356 DTSTAMP:19700101T120017Z DESCRIPTION:On Perfectly Secure Two-Party Computation for Symmetric Functionalities with Correlated Randomness URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/356/on-perfectly-secure-two-party-computation-for-symmetric-functionalities-with-correlated-randomness/ SUMMARY:A multiparty computation protocol is perfectly secure for some function f if it perfectly emulates an ideal computation of f. Thus, perfect security is the strongest and most desirable notion of security, as it guarantees security in the face of any adversary and eliminates the dependency on any security parameter. Ben-Or et al. [STOC 88] and Chaum et al. [STOC 88] showed that any function can be computed with perfect security if strictly less than one-third of the parties can be corrupted. For two-party sender-receiver functionalities (where only one party receives an output), Ishai et al. [TCC 13] showed that any function can be computed with perfect security in the correlated randomness model. Unfortunately, they also showed that perfect security cannot be achieved in general for two-party functions that give outputs to both parties (even in the correlated randomness model). We study the feasibility of obtaining perfect security for deterministic symmetric two-party functionalities (i.e., where both parties obtain the same output) in the face of malicious adversaries. We explore both the plain model as well as the correlated randomness model. We provide positive results in the plain model, and negative results in the correlated randomness model. As a corollary, we obtain the following results. - We provide a characterization of symmetric functionalities with (up to) four possible outputs that can be computed with perfect security. The characterization is further refined when restricted to three possible outputs and to Boolean functions. All characterizations are the same for both the plain model and the correlated randomness model. -We show that if a functionality contains an embedded XOR or an embedded AND, then it cannot be computed with perfect security (even in the correlated randomness model). DTSTART:20221124T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221201T120000Z UID:ec9a45f72601dbd22fea4ca69bc7745f-357 DTSTAMP:19700101T120017Z DESCRIPTION:Verifiable Relation Sharing and Multi-Verifier Zero-Knowledge in Two Rounds: Trading NIZKs with Honest Majority URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/357/verifiable-relation-sharing-and-multi-verifier-zero-knowledge-in-two-rounds-trading-nizks-with-honest-majority/ SUMMARY:We introduce the problem of Verifiable Relation Sharing (VRS) where a client (prover) wishes to share a vector of secret data items among k servers (the verifiers) while proving in zero knowledge that the shared data satisfies some properties. This combined task of sharing and proving generalizes notions like verifiable secret sharing and zero-knowledge proofs over secret-shared data. We study VRS from a theoretical perspective and focus on its round complexity. As our main contribution, we show that every efficiently-computable relation can be realized by a VRS with an optimal round complexity of two rounds where the first round is input-independent (offline round). The protocol achieves full UC-security against an active adversary that is allowed to corrupt any t-subset of the parties that may include the client together with some of the verifiers. For a small (logarithmic) number of parties, we achieve an optimal resiliency threshold of t < 0.5(k + 1), and for a large (polynomial) number of parties, we achieve an almost-optimal resiliency threshold of t < 0.5(k + 1)(1 − ε) for an arbitrarily small constant ε > 0. Both protocols can be based on sub-exponentially hard injective one-way functions. If the parties have an access to a collision resistance hash function, we can derive statistical everlasting security, i.e., the protocols are secure against adversaries that are computationally bounded during the protocol execution and become computationally unbounded after the protocol execution. Previous 2-round solutions achieve smaller resiliency thresholds and weaker security notions regardless of the underlying assumptions. As a special case, our protocols give rise to 2-round offline/online constructions of multi-verifier zero-knowledge proofs (MVZK). Such constructions were previously obtained under the same type of assumptions that are needed for NIZK, i.e., public-key assumptions or random-oracle type assumptions (Abe et al., Asiacrypt 2002; Groth and Ostrovsky, Crypto 2007; Boneh et al., Crypto 2019; Yang, and Wang, Eprint 2022). Our work shows, for the first time, that in the presence of an honest majority these assumptions can be replaced with more conservative “Minicrypt”-type assumptions like injective one-way functions and collision-resistance hash functions. Indeed, our MVZK protocols provide a round-efficient substitute for NIZK in settings where honest-majority is present. This is a joint work with Benny Applebaum and Arpita Patra. DTSTART:20221201T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221117T120000Z UID:9947b9637c81b3af3f08a11097d97539-358 DTSTAMP:19700101T120011Z DESCRIPTION:The Role of Adaptivity in Learning and Decision-Making URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/358/the-role-of-adaptivity-in-learning-and-decision-making/ SUMMARY:In many machine learning applications the learner is faced with a *stochastic environment* and it (sequentially) probes or influences the environment so as to optimize a given objective function. Examples of such applications include recommendation systems, web advertising, viral marketing, clinical trials, search ranking etc. For instance, in recommendation systems, the learner attempts to identify good recommendations by probing the stochastic preferences of users. Similarly, in viral marketing, the learner attempts to spread information through a social network using marketing campaigns that influence (stochastic) subsets of the network. Most existing learning algorithms for these applications operate in one of two settings: (1) non-adaptive, and (2) fully adaptive. In the non-adaptive setting, all the selections/probes are completely determined ahead of time. However, these a priori selections might be inefficient as some of them might be unnecessary in hindsight. In the fully adaptive setting, the selection policy is updated after each observation from the environment. However, this fully adaptive setting might not be practical in many applications due to delays in receiving observations from many parallel sources. In this talk we introduce a semi-adaptive setting that interpolates between these two extreme settings for a wide range of learning and decision-making problems such as best arm identification in multi-armed bandits, ranking from pairwise comparisons, dueling bandits, and stochastic submodular maximization. We show that semi-adaptive policies enjoy the power of fully adaptive policies while requiring very few updates to the selection/probing rules. We also identify the trade-offs between rounds of adaptivity and performance. DTSTART:20221117T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221117T120000Z UID:ecd022e50c9e5595c3f58cdf147a5c3c-359 DTSTAMP:19700101T120016Z DESCRIPTION:Birthday Paradox, Monochromatic Subgraphs, and Algorithmic Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/359/birthday-paradox-monochromatic-subgraphs-and-algorithmic-applications/ SUMMARY:What is the chance that among a group of friends, there are friends all of whom have the same birthday? This is the celebrated birthday problem which can be formulated as the existence of a monochromatic -clique (matching birthdays) in the complete graph , where every vertex of is uniformly colored with 365 colors (corresponding to birthdays). More generally, for a connected graph, let be the number of monochromatic copies of in a uniformly random coloring of the vertices of the graph with colors. In this talk, the asymptotic properties of this quantity will be derived, and applications in distributional property testing and computation of discrete logarithms will be discussed. DTSTART:20221117T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221124T120000Z UID:6c4fefe7e48088475709164068b2585b-360 DTSTAMP:19700101T120015Z DESCRIPTION:Doing your Lifes work URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/360/doing-your-lifes-work/ SUMMARY:The rapid evolution of technologies within and outside a company deeply affect the choice of work we intend to do as our lifes work. As a case study, in this talk, I will give a high-level overview of technologies developed in NVIDIA, and the backdrop in which these technologies evolved. Further, I will give a birds eye view of the implications of the dynamic settings in the current day workplace to doing our lifes work. DTSTART:20221124T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221125T120000Z UID:366236faa1cb65b315bcd3cfacc67430-361 DTSTAMP:19700101T120011Z DESCRIPTION:Multi-Party Computing for Privacy in Machine Learning Systems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/361/multi-party-computing-for-privacy-in-machine-learning-systems/ SUMMARY:Given the resource management benefits such as elasticity, availability, and cost-effectiveness offered by cloud service providers, a growing number of machine learning workloads are migrated to the cloud for operations. Under this modern compute paradigm, confidential data and models can be leaked to unwanted parties if the service providers are curious, malicious, or compromised. The privacy concern is particularly pressing for natural language processing (NLP) where user’s audio features are inputs to ML models. These inputs contain sensitive private information about the users and require rigorous protection. Secure multi party computing (MPC) is one approach to tackle the privacy leaks without relying on any additional hardware support. MPC protocols provide strong security even when a subset of parties are compromised. However, when it comes to protecting privacy there is no free lunch, and in fact we show that it is a very expensive lunch. Through a detailed characterization of industry-strength MPC implementation of Transformer-based NLP models, we analyze where the MPC performance bottlenecks are. First, we show that Transformers rely extensively on softmax <br> Talk link <a href="https://youtu.be/2nE6nbkfuls">https://youtu.be/2nE6nbkfuls</a> <br> DTSTART:20221125T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221125T120000Z UID:d30e7b289902a916b201ca047e84b75e-362 DTSTAMP:19700101T120011Z DESCRIPTION:Comparative Analysis of Topological Structures URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/362/comparative-analysis-of-topological-structures/ SUMMARY:Measuring scientific processes result in a set of scalar functions (scalar fields) which may be related temporally, be part of an ensemble, or unrelated. Overall understanding and visualization of scientific processes require the study of individual fields and, more importantly, the development of methods to compare them meaningfully. In this thesis, we focus on the problem of designing meaningful measures to compare scalar fields by comparing their abstract representations called topological structures. We emphasize on intuitive and practical measures with useful properties and applications. The first part of the thesis deals with comparing a topological structure called the merge tree. We propose two global comparison measures, both based on tree edit distances. The first measure OTED is based on the assumption that merge trees are ordered rooted trees. Upon finding that there is no meaningful way of imposing such an order, we propose a second measure called MTED for comparing unordered rooted trees. We propose intuitive cost models and prove that MTED is a metric. We also provide various applications such as shape comparison, periodicity detection, symmetry detection, temporal summarization, and an analysis of the effects of sub-sampling /smoothing on the topology of the scalar field. The second part deals with a local comparison measure LMTED for merge trees that supports the comparison of substructures of scalar fields, thus facilitating hierarchical or multi-scale analysis and alleviating some drawbacks of MTED. We propose a dynamic programming algorithm, prove that LMTED is a metric and also provide applications such as symmetry detection in multiple scales, a finer level analysis of sub-sampling effects, an analysis of the effects of topological compression, and feature tracking in time-varying fields. The third part of the thesis deals with comparison of a topological structure called the extremum graph. We provide two comparison measures for extremum graphs based on persistence distortion (PDEG) and Gromov-Wasserstein distance (GWEG). Both persistence distortion and Wasserstein distance are known metrics. We analyze how the underlying metric affects these comparison measures and present various applications such as periodicity detection to facilitate scientific data analysis and visualization. The final part of the thesis introduces a time-varying version of extremum graphs (TVEG) with a simple comparison criterion to identify correspondences between features in successive time steps. We provide applications to tracking features in time-varying scalar fields from computational fluid dynamics. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_YWQ3OGE4ODEtYmJiNi00OGMxLWFjMTEtN2ZkMzI0YTYxYzVi%40thread.v2/0?context={ DTSTART:20221125T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221128T120000Z UID:36e7ea0918e412e676290cba332b9e31-363 DTSTAMP:19700101T120011Z DESCRIPTION:Approximate Representation of Symmetric Submodular Functions via Hypergraph Cut Functions URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/363/approximate-representation-of-symmetric-submodular-functions-via-hypergraph-cut-functions/ SUMMARY:Submodular functions are fundamental to combinatorial optimization. We consider the problem of approximating symmetric submodular functions everywhere using hypergraph cut functions. Prior works have shown that symmetric submodular functions over n-element ground sets cannot be approximated within a (n/8)-factor using a graph cut function and raised the question of approximating them using hypergraph cut functions. In this talk, I will show that there exist symmetric submodular functions over n-element ground sets that cannot be approximated within a o(n^{1/3}/log^2 n)-factor using a hypergraph cut function. On the positive side, I will discuss the approximability of symmetrized concave linear functions and symmetrized rank functions of uniform matroids and partition matroids using hypergraph cut functions. Organizers Note:: Based on joint work with Calvin Beideman, Chandra Chekuri, and Chao Xu. Speaker Website: http://karthik.ise.illinois.edu/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d For more details please visit: https://www.csa.iisc.ac.in/iisc-msr-seminar/ Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20221128T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221128T120000Z UID:821a22ac8dc68f044999f39a2e580bad-364 DTSTAMP:19700101T120011Z DESCRIPTION:Approximate Representation of Symmetric Submodular Functions via Hypergraph Cut Functions URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/364/approximate-representation-of-symmetric-submodular-functions-via-hypergraph-cut-functions/ SUMMARY:Submodular functions are fundamental to combinatorial optimization. We consider the problem of approximating symmetric submodular functions everywhere using hypergraph cut functions. Prior works have shown that symmetric submodular functions over n-element ground sets cannot be approximated within a (n/8)-factor using a graph cut function and raised the question of approximating them using hypergraph cut functions. In this talk, I will show that there exist symmetric submodular functions over n-element ground sets that cannot be approximated within a o(n^{1/3}/log^2 n)-factor using a hypergraph cut function. On the positive side, I will discuss the approximability of symmetrized concave linear functions and symmetrized rank functions of uniform matroids and partition matroids using hypergraph cut functions. Based on joint work with Calvin Beideman, Chandra Chekuri, and Chao Xu. Speaker Website: http://karthik.ise.illinois.edu/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d For more details please visit: https://csa.iisc.ac.in/theoryseminars/ Hosts: Rahul Madhavan and Rameesh Paul DTSTART:20221128T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221216T120000Z UID:e6a164c275268dd7789216028ab200c5-365 DTSTAMP:19700101T120015Z DESCRIPTION:Anti-virus hardware: Applications in Embedded, Automotive and Power Systems security. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/365/anti-virus-hardware-applications-in-embedded-automotive-and-power-systems-security/ SUMMARY:Anti-virus software (AVS) tools are used to detect Malware in a system. However, software-based AVS are vulnerable to attacks. A malicious entity can exploit these vulnerabilities to subvert the AVS. Recently, hardware components such as Hardware Performance Counters (HPC) have been used for Malware detection, in the form of Anti-virus Hardware (AVH). In this talk, we will discuss HPC-based AVHs for improving embedded security and privacy. Furthermore, we will discuss the application of HPCs in security cyber physical systems (CPS), namely automotive and microgrid systems. Subsequently, we will discuss their pitfalls. Finally, we will present PREEMPT, a zero overhead, high-accuracy and low-latency technique to detect Malware by re-purposing the embedded trace buffer (ETB), a debug hardware component available in most modern processors. The ETB is used for post-silicon validation and debug and allows us to control and monitor the internal activities of a chip, beyond what is provided by the Input/Output pins. PREEMPT combines these hardware-level observations with machine learning-based classifiers to preempt Malware before it can cause damage. We will conclude the talk with future research directions and challenges. DTSTART:20221216T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221202T120000Z UID:68f22a8f97f8fbb19a2be4b229a92690-366 DTSTAMP:19700101T120011Z DESCRIPTION:Robustly Learning Mixtures of Arbitrary Gaussians in Polynomial Time URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/366/robustly-learning-mixtures-of-arbitrary-gaussians-in-polynomial-time/ SUMMARY:The Gaussian Mixture Model (Pearson 1894) is widely used for high-dimensional data. While classical results establish its unique identifiability, it was shown in 2010 (Kalai-Moitra-Valiant, Belkin-Sinha) that for any fixed number of component Gaussians, the underlying mixture parameters can be estimated to arbitrary accuracy in polynomial time. Robust Statistics (Huber 1964) asks for estimation of underlying models robustly, i.e., even if a bounded fraction of data is noisy, possibly chosen adversarially. This goal seemed to be computationally intractable, even for estimating the mean of nice distributions, till 2016 (Diakonikolas-Kamath-Kane-Li-Moitra-Stewart, Lai-Rao-Vempala). These methods were extended to many problems, but the robust estimation of GMMs remained a central open problem. In this talk, we will present the first polytime algorithm for any fixed number of components with no assumptions on the underlying mixture. The techniques developed, clustering using convex relaxations and approximate tensor decomposition allowing for error in both Frobenius norm and low-rank terms, might be useful more generally. Speaker Website https://faculty.cc.gatech.edu/~vempala/ Microsoft teams link: <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> Hosts: Aditya Abhay Lonkar, Rahul Madhavan and Rameesh Paul DTSTART:20221202T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221205T120000Z UID:e561370827450e4a956ce0a7cc48d209-367 DTSTAMP:19700101T120011Z DESCRIPTION:Exploring the Gap between Tolerant and Non-tolerant Distribution Testing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/367/exploring-the-gap-between-tolerant-and-non-tolerant-distribution-testing/ SUMMARY:The framework of distribution testing is currently ubiquitous in the field of property testing. In this model, the input is a probability distribution accessible via independently drawn samples from an oracle. The testing task is to distinguish a distribution that satisfies some property from a distribution that is far in some distance measure from satisfying it. The task of tolerant testing imposes a further restriction, that distributions close to satisfying the property are also accepted. This work focuses on the connection between the sample complexities of non-tolerant testing of distributions and their tolerant testing counterparts. When limiting our scope to label-invariant (symmetric) properties of distributions, we prove that the gap is at most quadratic, ignoring poly-logarithmic factors. Conversely, the property of being the uniform distribution is indeed known to have an almost-quadratic gap. Moreover, we prove lower bounds on the sample complexities of non-tolerant as well as tolerant testing for a special class of distribution properties, namely non-concentrated distribution properties, where the probability mass of the distributions in the property is sufficiently spread out. Finally, we design an algorithm that can learn a concentrated distribution even when its support set is unknown apriori. This is a joint work with Sourav Chakraborty, Eldar Fischer, Arijit Ghosh and Gopinath Mishra. Speaker Website https://sites.google.com/view/sayantans/home Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Aditya Abhay Lonkar, Rahul Madhavan and Rameesh Paul DTSTART:20221205T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221215T120000Z UID:627e97d29230d2d1ba3d5e580369f41f-368 DTSTAMP:19700101T120016Z DESCRIPTION:Fairness in AI-based Decision Making URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/368/fairness-in-ai-based-decision-making/ SUMMARY:AI systems are ubiquitous in the current times, facilitating numerous real-world and even real-time applications. The existing models achieve near-optimal results for specific performance measures. Such perfection is often obtained at the cost of DTSTART:20221215T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221214T120000Z UID:00b9a6273bef7d0b38b6636e8ab28bd4-369 DTSTAMP:19700101T120015Z DESCRIPTION:Applications of Dedekinds Index Theorem to Lattice-based Cryptography URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/369/applications-of-dedekinds-index-theorem-to-lattice-based-cryptography/ SUMMARY:Computationally hard problems on integer lattices, such as shortest vector problem (SVP), have become an important tool in designing modern cryptographic schemes, especially since these problems are considered post-quantum secure. For example, Shors quantum polynomial-time algorithm for integer factorization has rendered the famous RSA cryptosystem insecure, assuming existence of quantum computers. Instead of basing hardness on worst-case integer lattices, to make the lattice-based encryption schemes more efficient, there has been a significant push to use ideal lattices in ring of integers of number fields, and such a scheme has even been standardized by NIST recently. However, it is not clear if the additional algebraic structure of such ideal lattices, which lie in well known Dedekind-domains, can withstand quantum attacks. In this work we show that we can base similar and natural cryptosystems on hardness of ideal lattices in non Dedekind-domains which have less algebraic structure. This allows the security of the efficient cryptosystems to be based on problems closer to the worst-case integer lattices. The main technical contribution of the work is a novel and generalized way to prove the “ideal-clearing lemma” of Lyubashevsky et al (Eurocrypt 2010). This is joint work with Chengyu Lin. DTSTART:20221214T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221212T120000Z UID:3e573f1844aebce23555c8e212f51a67-371 DTSTAMP:19700101T120016Z DESCRIPTION:An MLIR-based High-level Synthesis Compiler for Hardware Accelerator Design URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/371/an-mlir-based-high-level-synthesis-compiler-for-hardware-accelerator-design/ SUMMARY:The emergence of machine learning, image and audio processing on edge devices has motivated research towards power-efficient custom hardware accelerators. Though FPGAs are an ideal target for custom accelerators, the difficulty of hardware design and the lack of vendor-agnostic, standardized hardware compilation infrastructure has hindered their adoption. High-level synthesis (HLS) offers a more compiler-centric alternative to the traditional Verilog-based hardware design improving developer productivity. In this work, we propose an MLIR-based end-to-end HLS compiler and an an intermediate representation that is suitable for the design and implementation of domain-specific accelerators for affine workloads. Our compiler brings similar levels of modularity and extensibility to the HLS compilation domain, which LLVM brought to the area of a software compilation. A modular compiler infrastructure offers the advantage of incrementally introducing new language frontends and optimization passes without the need to reinvent the whole HLS compiler stack. Our compiler converts a high-level description of the accelerator specified in the C programming language into a register-transfer-level design in SystemVerilog. We use memory dependence analysis and integer-linear-program(ILP) based automatic scheduling on improving loop-pipelining and introduce parallelization between producer-consumer kernels. Our ILP-based optimizer beats the state-of-the-art Vitis HLS compiler by 1.3x in performance over a representative set of benchmarks while requiring fewer FPGA resources. Microsoft Teams meeting Join on your computer, mobile app or room device https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjZiOWVkMzgtOGU2MS00YjExLWE1ZTctNjBkNmYzZjQwODE4%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22171d9abc-cf43-429a-9680-c05b9523fa9a%22%7d Meeting ID: 448 946 611 293 Passcode: gQbJz4 DTSTART:20221212T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221212T120000Z UID:0d63d38fa422805402890ef5a453308b-373 DTSTAMP:19700101T120012Z DESCRIPTION:Stochastic Optimization And Its Application In Reinforcement Learning. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/373/stochastic-optimization-and-its-application-in-reinforcement-learning/ SUMMARY:Numerous engineering fields, including transportation systems, manufacturing, communication networks, healthcare, and finance, frequently encounter problems requiring optimization, including uncertainty. Simulation-based optimization is a workable substitute for accurate analytical solutions because of the numerous input variables and the need for a system model. Smoothed functional (SF) algorithms belong to the class of simultaneous perturbation methods that have been found useful for stochastic optimization problems particularly in high-dimensional parameter spaces. SF methods update the gradient of the objective using function measurements involving parameters that are perturbed simultaneously along all component directions. Katkovnik and Kulchitsky originally developed the SF gradient procedure. This results in the objective function getting smoothed because of the convolution. The objective function smoothing that results from the convolution with a smoothing density function can help the algorithm converge to a global minimum or to a point close to it. First we present a stochastic gradient algorithm for minimizing a smooth objective function that is an expectation over noisy cost samples, and only the latter are observed for any given parameter. Our algorithm employs a gradient estimation scheme with random perturbations, which are formed using the truncated Cauchy distribution from the $delta$ sphere. We analyze the bias and variance of the proposed gradient estimator. Our algorithm is found to be particularly useful in the case when the objective function is non-convex, and the parameter dimension is high. From an asymptotic convergence analysis, we establish that our algorithm converges almost surely to the set of stationary points of the objective function and obtain the asymptotic convergence rate. We also show that our algorithm avoids unstable equilibria, implying convergence to local minima. Further, we perform a non-asymptotic convergence analysis of our algorithm. In particular, we establish here a non-asymptotic bound for finding an $epsilon$-stationary point of the non-convex objective function. Finally, we demonstrate numerically through simulations that our algorithm outperforms GSF, SPSA and RDSA by a significant margin over a few non-convex settings and we further validate its performance over convex (noisy) objectives. Next we consider the problem of control in the setting of reinforcement learning (RL), where model information is not available. Policy gradient algorithms are a popular solution approach for this problem, and are usually shown to converge to a stationary point of the value function. We propose two policy Newton algorithms that incorporate cubic regularization. Both algorithms employ the likelihood ratio method to form estimates of the gradient and Hessian of the value function using sample trajectories. The first algorithm requires exact solution of the cubic regularized problem in each iteration, while the second algorithm employs an efficient gradient descent-based approximation to the cubic regularized problem. We establish convergence of our proposed algorithms to a second-order stationary point (SOSP) of the value function, which results in avoidance of traps in the form of saddle points. In particular, the sample complexity of our algorithms towards finding an $epsilon$-SOSP is $O(epsilon^{-3.5})$, and this is a significant improvement over the state-of-the-art sample complexity of $O(epsilon^{-4.5})$. Microsoft teams link: Join on your computer, mobile app or room device https://teams.microsoft.com/l/meetup-join/19%3ameeting_OGI4YjEzMmUtNGFkZi00YmE5LTgzMzQtYWVmZTI5OGQxNTE3%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22578cbfbe-b1e9-4f20-91bd-984b656368e5%22%7d Meeting ID: 488 618 948 051 Passcode: XQantJ DTSTART:20221212T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221209T120000Z UID:98fe5aa4e2bc4e954b7d86912e9fb993-374 DTSTAMP:19700101T120016Z DESCRIPTION:Static Race Detection for Periodic Real-Time Programs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/374/static-race-detection-for-periodic-real-time-programs/ SUMMARY:We consider the problem of statically detecting data races in periodic real-time programs that use locks, and run on a single processor platform. We propose a technique based on a small set of rules that exploits the priority, periodicity, locking, and timing information of tasks in the program. One of the key requirements is a response time analysis for such programs, and we propose an algorithm to compute this for the case of non-nested locks. We have implemented our analysis for real-time C programs in a tool called PePRacer and evaluated its performance on a small set of benchmarks from the literature. This is joint work with Varsha Suresh (IIITB), Rekha Pai (IISc/Oxford), Sujit Kumar Chakrabarty (IIITB) and Meenakshi DSouza (IIITB). DTSTART:20221209T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221219T120000Z UID:92fbdc54229de092a104d0da812783bb-375 DTSTAMP:19700101T120011Z DESCRIPTION:Unified Question Answering over RDF Knowledge Graphs and Natural Language Text URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/375/unified-question-answering-over-rdf-knowledge-graphs-and-natural-language-text/ SUMMARY:Question answering over knowledge graphs and other RDF data has been greatly advanced, with a number of good systems providing crisp answers for natural language questions or telegraphic queries. Some of these systems incorporate textual sources as additional evidence for the answering process, but cannot compute answers that are present in text alone. Conversely, systems from the IR and NLP communities have addressed QA over text, but such systems barely utilize semantic data and knowledge. In this work, we develop a QA system that can seamlessly operate over RDF datasets and text corpora, or both together, in a unified framework. Our method, called Uniqorn, builds a context graph on-the-fly, by retrieving question-relevant triples from the RDF data and/or snippets from a text corpus, using a fine-tuned BERT model. The resulting graph is typically rich but highly noisy. Uniqorn copes with this input by advanced graph algorithms for Group Steiner Trees, that identify the best answer candidates in the context graph. Experimental results on several benchmarks of complex questions with multiple entities and relations, show that Uniqorn produces results comparable to the state-of-the-art on KGs, text corpora, and heterogeneous sources. The graph-based methodology provides user-interpretable evidence for the complete answering process. DTSTART:20221219T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221209T120000Z UID:385fcd0caa11b728893005ceeefe199b-376 DTSTAMP:19700101T120014Z DESCRIPTION:Lifelong Learning of Representations with Provable Guarantees URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/376/lifelong-learning-of-representations-with-provable-guarantees/ SUMMARY:In lifelong learning, tasks (or classes) to be learned arrive sequentially over time in arbitrary order. During training, knowledge from previous tasks can be captured and transferred to subsequent ones to improve sample efficiency. We consider the setting where all target tasks can be represented in the span of a small number of unknown linear (or nonlinear) features of the input data and propose a lifelong learning algorithm that maintains and refines the internal feature representation. We prove that for any desired accuracy on all tasks, the dimension of the representation remains close to that of the underlying representation. The resulting algorithm is provably efficient and the sample complexity for input dimension d, m tasks with k total features up to error ϵ is O~((dk^1.5 + km)/ϵ). We also prove a matching lower bound for any lifelong learning algorithm that uses a single task learner as a black box. An empirical study, with a lifelong learning heuristic for deep neural networks, performs favorably on challenging image datasets compared to state-of-the-art continual learning methods. Speaker Website https://faculty.cc.gatech.edu/~vempala/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Aditya Abhay Lonkar, Rahul Madhavan, Rameesh Paul & Aditya Subramanian DTSTART:20221209T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221215T120000Z UID:4f4f716c4c100e13c9e78a612522d63d-377 DTSTAMP:19700101T120011Z DESCRIPTION:HYDRA: Dynamic Approach to Database Regeneration URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/377/hydra-dynamic-approach-to-database-regeneration/ SUMMARY:Database software vendors often need to generate synthetic databases for a variety of applications, including (a) Testing database engines and applications, (b) Data masking, (c) Benchmarking, (d) Creating what-if scenarios, and (e) Assessing performance impacts of planned engine upgrades. The synthetic databases are targeted toward capturing the desired schematic properties (e.g., keys, referential constraints, functional dependencies, domain constraints), as well as the statistical data profiles (e.g., value distributions, column correlations, data skew, output volumes) hosted on these schemas. Several data generation frameworks have been proposed in the last two decades. It started from the ab-initio generation tools that use standard mathematical distributions and do not depend on the client databases or query workloads. Subsequently, tools that generate data using column distributions became prominent. However, none of these mechanisms could mimic the customer query-processing environments satisfactorily. The contemporary school of thought generates workload-aware data that uses query execution plans from the customer workloads as input and guarantees volumetric similarity. That is, the intermediate row cardinalities obtained at the client and vendor sites are very similar when matching query plans are executed. This similarity helps to preserve the multi-dimensional layout and flow of the data, a prerequisite for achieving similar performance on the clients workload. However, even in this category, the existing frameworks are crippled by one or more of the limitations on several fronts, such as the inability to (a) provide a comprehensive algorithm to handle the queries based on core relational algebra operators, namely, select, project, and join; (b) scale to big data volumes; (c) scale to large input workloads; and (d) provide high accuracy on unseen queries. In this work, motivated by the above lacuna, we present HYDRA, a data regeneration tool that materially addresses the above challenges by adding functionality, dynamism, scale, and robustness. Firstly, the extended workload coverage is obtained by providing a comprehensive solution to support queries based on select-project-join relational algebra operators. Specifically, the constraints are modeled using a linear feasibility problem, in which each variable represents the volume of a region of the data space. These regions are computed using a scheme of partitioning strategies. For example, to encode the filter constraints, our region-partitioning approach divides the data space into the provably minimum number of regions, thereby reducing the existing solutions complexity by many orders of magnitude. Our projection subspace division and projection isolation strategies address the critical challenges in incorporating projection-inclusive constraints. By modeling referential constraints over denormalized equivalents of the tables, Hydra delivers a comprehensive solution that also additionally handles join constraints. Secondly, a unique feature of our data regeneration approach is that it delivers a database summary as the output rather than the static data itself. This summary is of negligible size and depends only on the query workload and not on the database scale. It can be used for dynamically generating data during query execution. Therefore, the enormous time and space overheads incurred by prior techniques in generating and storing the data before initiating analysis are eliminated. Specifically, the summaries for complex Big Data client scenarios comprising over a hundred queries are constructed within just a few minutes, requiring only a few MBs of storage. We have evaluated the proposed ideas using both synthetic benchmarks, such as TPC-DS, and real-world benchmarks based on Census and IMDB databases. Thirdly, to improve accuracy towards unseen queries, Hydra additionally exploits metadata statistics maintained by the database engine. Specifically, it adds an objective function to the linear program to pick a solution with improved inter-region tuple distribution. Further, a uniform distribution of tuples within regions is generated to get a spread of values. In a nutshell, these techniques facilitate the careful selection of a desirable database from the candidate synthetic databases and also provide metadata compliance. Lastly, as a proof of concept, the Hydra framework has been prototyped in a Java-based tool that provides a visual and interactive demonstration of the data regeneration pipeline. The tool has been warmly received by both academic and industrial communities. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWU3MGY2NzItZDJmZS00ZjYzLWI1MTQtNjc2MjExMWI5ZWJk%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22408cb270-9022-405b-b737-6c9e2ec3fb15%22%7d DTSTART:20221215T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221220T120000Z UID:eb3c199ce50a9d7158757ad1e5acc2fb-378 DTSTAMP:19700101T120015Z DESCRIPTION:Planet Scale Computing Infrastructure - Challenges and Opportunities URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/378/planet-scale-computing-infrastructure-challenges-and-opportunities/ SUMMARY:There are at least a dozen applications from Google that are used by over a billion users worldwide regularly. We will discuss a sketch of how a Planet Scale Computing Infrastructure was architected and designed by Google to enable billions of users worldwide seamless access to these services. We will discuss the organizational sketch, and what we have learned from the experience of developing a Planet Scale Computing infrastructure. We will outline the key challenges and opportunities we see going forward that will need critical and broad innovation to ensure that an affordable computing infrastructure continues to be available to users. We will make the case that addressing these challenges to enable the large class of emerging applications and disruptions that we envision in an AI/ML driven technology landscape is the need of the hour. DTSTART:20221220T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221216T120000Z UID:42a718ed3cc3f0439d9e67d4e63dedc4-379 DTSTAMP:19700101T120016Z DESCRIPTION:Enabling Efficient Memory Systems using Novel Compression Methods URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/379/enabling-efficient-memory-systems-using-novel-compression-methods/ SUMMARY:Using data compression methods in the memory hierarchy can improve the efficiency of memory systems by enabling higher effective cache capacity, more effective use of available memory bandwidth and by enabling higher effective main memory capacity. This can lead to higher substantially higher performance and lower power consumption. However, to enable these values requires highly effective compression algorithms that can be implemented with low latency and high throughput. Research at Chalmers University of Technology and at ZeroPoint Technologies, a fabless startup company, has yielded many new families of compression methods that are now being commercially deployed. This talk will present the major insights of more than a decade of research on memory compression methods for the memory hierarchy. The talk covers value-aware and statistical compression caches, compression algorithms that are tuned to the data at hand through data analysis using new clustering algorithms to allow for substantially higher memory bandwidth. DTSTART:20221216T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221221T120000Z UID:d23d8b6c88686d1634eb29b76f8eacfa-380 DTSTAMP:19700101T120010Z DESCRIPTION:Solving Global Grand Challenges with High Performance Data Analytics URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/380/solving-global-grand-challenges-with-high-performance-data-analytics/ SUMMARY:Data science aims to solve grand global challenges such as: detecting and preventing disease in human populations; revealing community structure in large social networks; protecting our elections from cyber-threats, and improving the resilience of the electric power grid. Unlike traditional applications in computational science and engineering, solving these social problems at scale often raises new challenges because of the sparsity and lack of locality in the data, the need for research on scalable algorithms and architectures, and development of frameworks for solving these real-world problems on high performance computers, and for improved models that capture the noise and bias inherent in the torrential data streams. In this talk, Bader will discuss the opportunities and challenges in massive data science for applications in social sciences, physical sciences, and engineering. DTSTART:20221221T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221223T120000Z UID:c57a0de88bf4a8b12dce774334109baf-381 DTSTAMP:19700101T120016Z DESCRIPTION:Matroid-convex functions and approximative closure of some polynomial classes. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/381/matroid-convex-functions-and-approximative-closure-of-some-polynomial-classes/ SUMMARY:Matroid-convex (or M-convex) functions, defined by Murota, are functions on size k subsets of a ground set which satisfy a Steinitzs exchange type inequality. M-convex function minimization can be done by some greedy-type algorithms. More interestingly, Murota proved a splitting theorem that reduces minimization of sum of two M-convex functions to minimizing two M-convex functions separately. An interesting question in algebraic complexity is to understand approximative closure of various classes of polynomials. We use the above splitting theorem to show that the following class of polynomials is closed under approximation: det(sum_i A_i x_i) where A_is are all rank 1 matrices. That is, any polynomial which can be approximately computed by this model can also be exactly computed by it. Based on a joint work with Abhranil Chatterjee, Sumanta Ghosh, and Roshan Raj Speaker Website https://www.cse.iitb.ac.in/~rgurjar/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Hosts: Aditya Subramanian, Aditya Abhay Lonkar, Rahul Madhavan & Rameesh Paul DTSTART:20221223T120000Z END:VEVENT BEGIN:VEVENT DTEND:20221223T120000Z UID:e65505db4730c759c5f2c45f48403fc7-382 DTSTAMP:19700101T120017Z DESCRIPTION:Efficient Determinant Maximization for All Matroids URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/382/efficient-determinant-maximization-for-all-matroids/ SUMMARY:Determinant maximization provides an elegant generalization of problems in many areas, including convex geometry, statistics, machine learning, fair allocation of goods, and network design. In an instance of the determinant maximization problem, we are given a collection of vectors $v_1,ldots, v_n in R^d$, and the goal is to pick a subset $Ssubseteq [n]$ of given vectors to maximize the determinant of the matrix $sum_{i in S} v_iv_i^top$, where the picked set of vectors $S$ must satisfy some combinatorial constraint such as cardinality constraint ($|S| leq k$) or matroid constraint ($S$ is a basis of a matroid defined on $[n]$). We give efficient deterministic combinatorial algorithms for the determinant maximization problem under a matroid constraint that achieves $O(r^{O(r)})$-approximation for any matroid of rank $rleq d$ and $O(d^O(d))$-approximation for any matroid of rank $rgeq d$. The algorithm for the $rleq d$ case relies on the geometric interpretation of the determinant whereas the algorithm for the $rgeq d$ case relies on the algebraic properties of the determinant and the properties of a convex programming relaxation introduced by Madan et al. (FOCS 20). In both cases, we use matroid intersection as a local search tool to iteratively improve a solution by finding an alternating negative cycle in an appropriately defined exchange graph defined by the matroids. Joint work with Adam Brown, Aditi Laddha, Mohit Singh, and Prasad Tetali. DTSTART:20221223T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230105T120000Z UID:7ef4a09db3444681ffb4de305c6cc90f-383 DTSTAMP:19700101T120016Z DESCRIPTION:Probabilistic Hash Functions and Hash Tables: A New Paradigm for Efficient AI Training and Inference URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/383/probabilistic-hash-functions-and-hash-tables-a-new-paradigm-for-efficient-ai-training-and-inference/ SUMMARY:Neural Scaling Law informally states that an increase in model size and data automatically improves AI. However, we have reached a point where the growth has reached a tipping end where the cost and energy associated with AI are becoming prohibitive. This talk will demonstrate the algorithmic progress that can exponentially reduce the compute and memory cost of training and inference with neural networks. We will show how data structures can fundamentally break the barriers of some of the classical adaptive sampling subroutines. In particular, randomized hash tables can be used to design an efficient DTSTART:20230105T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230105T120000Z UID:518044090a9929572cf7bbafe1d41ff8-384 DTSTAMP:19700101T120017Z DESCRIPTION:Civic Agency in AI: Rethinking Responsible Practices & Critical Discourses in Finland & the E.U. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/384/civic-agency-in-ai-rethinking-responsible-practices-critical-discourses-in-finland-the-e-u/ SUMMARY:The public sector is increasingly embracing algorithmic decision-making and data-centric infrastructures to improve digital services in areas such as education, healthcare, and urban mobility. Some AI-based systems are being used by governments for biometric surveillance, criminal justice, and other forms of citizen monitoring, which pose higher risks for abuse and unfair incrimination if they are not made easily transparent, accountable, or their legitimate use challenged by civil society. With regulations like the AI Act and the AI Liability Directive emerging in the EU, organizations must comply with complex ethical and regulatory frameworks. In our research we are exploring novel AI-based innovations, regulations, and practices, particularly the role of integrative software frameworks (MLOps and RegOps) and regulatory AI sandboxes, to facilitate experimentation, co-learning, and responsible deployment throughout the AI lifecycle. We are also examining wide-ranging public discourses around AI, using a mix of qualitative methods and Natural Language Processing (NLP), both among actors who influence its development and the publics affected by it. Linguistic devices such as metaphors, metonymy, and personification reveal how we conceptualize, narrate, contest, or attribute agency to AI systems. Our research can demonstrate how language affects attitudes, influences practices and policies, and shapes future imaginaries around AI. How must we reframe such narratives while fostering greater human responsibility and civic agency in AI systems, to make them more trustworthy, inclusive, and accountable in the future? My talk will draw on critical transdisciplinary perspectives and applied research in the Finnish and European context. DTSTART:20230105T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230112T120000Z UID:b154bbbb1e7f2f156c6b916814993be6-385 DTSTAMP:19700101T120011Z DESCRIPTION:Towards Next-Generation ML/AI: Robustness, Optimization, Privacy. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/385/towards-next-generation-ml-ai-robustness-optimization-privacy/ SUMMARY:Two trends have taken hold in machine learning and artificial intelligence: a move to massive, general purpose, pre-trained models as well as a move to small, on-device models trained on distributed data. Both these disparate settings face some common challenges: a need for (a) robustness to deployment conditions that differ from training, (b) faster optimization, and (c) protection of data privacy. As a result of the former trend, large language models have displayed emergent capabilities they have not been trained for. Recent models such as GPT-3 have attained the ability to generate remarkably human-like long-form text. I will describe Mauve, a measure to quantify the goodness of this emergent capability. It measures the gap between the distribution of generated text and that of human-written text. Experimentally, Mauve correlates the strongest with human evaluations of the generated text and can quantify a number of its qualitative properties. The move to massively distributed on-device federated learning of models opens up new challenges due to the natural diversity of the underlying user data and the need to protect its privacy. I will discuss how to reframe the learning problem to make the model robust to natural distribution shifts arising from deployment on diverse users who do not conform to the population trends. I will describe a distributed optimization algorithm and show how to implement it with end-to-end differential privacy. To conclude, I will discuss my ongoing efforts and future plans to work toward the next generation of ML/AI techniques by combining the best of both worlds. I will discuss applications ranging from differentially private language models and text generation to decentralized learning. DTSTART:20230112T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230105T120000Z UID:b9b06c7a68e7abbfe8ddf77b53c8c08d-386 DTSTAMP:19700101T120014Z DESCRIPTION:Quantum Worst-case to Average-case reductions for all linear problems. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/386/quantum-worst-case-to-average-case-reductions-for-all-linear-problems/ SUMMARY:Given an algorithm that has a small non-zero probability of answering correctly on an average input, can we use it to design another algorithm that has non-zero probability of answering correctly even on worst-case inputs? In this talk, I will focus on quantum algorithms for linear problems, and describe an explicit and efficient transformation that turns algorithms which are only correct on a small (even sub-constant) fraction of their inputs into ones that are correct on all inputs. This stands in contrast to the classical setting, where such results are only known for a small number of specific problems or restricted computational models. Along the way I will also present a tight Omega(n^2) lower bound on the average-case quantum query complexity of the Matrix-vector Multiplication problem. The techniques used in this work build on the recently introduced additive combinatorics framework for classical worst-case to average-case reductions (STOC 2022). The key quantum ingredients are subroutines based on quantum singular value transformations for approximate verification of the output of noisy quantum algorithms, and a learner for the heavy Fourier characters of indicator functions with imperfect quantum implementations. I will discuss how these tools can be combined to prove a quantum local correction lemma based on a probabilistic generalisation of Bogolyubovs lemma in additive combinatorics, leading to our worst-case to average-case transformation for linear problems. This talk is based on joint work with Vahid Asadi, Alexander Golovnev, Tom Gur, and Igor Shinkar (https://arxiv.org/abs/2212.03348). Speaker Website https://scholar.google.co.uk/citations?user=s02f_hIAAAAJ&hl=en Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Acknowledging the support from the Kirani family for generously supporting the seminar series. Hosts: Aditya Subramanian, Aditya Abhay Lonkar, Rahul Madhavan & Rameesh Paul DTSTART:20230105T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230109T120000Z UID:da1de9b6a4cbd880f360b71b837eb5f4-387 DTSTAMP:19700101T120011Z DESCRIPTION:Smoothed Analysis of Online Decision Making URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/387/smoothed-analysis-of-online-decision-making/ SUMMARY:We establish novel techniques to analyze algorithms in the smoothed analysis model for online decision making. In this model, at each step, an adversary chooses the input from distribution with density bounded above by the uniform distribution. Crucially, these techniques hold for adaptive adversaries that can choose distributions based on the decisions of the algorithm and the previous realizations of the inputs. Our technique effectively reduces the setting of adaptive adversaries to the simpler oblivious adversaries. The main application is to show that, in this model, online learning is as easy as offline learning. That is, we show that the regret against smoothed adversaries is captured by the offline complexity measure, VC dimension. Furthermore, we design efficient algorithms for online learning, circumventing impossibility results in the worst case. Based on joint works with Nika Haghtalab, Yanjun Han, Tim Roughgarden and Kunhe Yang. Speaker Website https://ashettyv.github.io/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Acknowledging the support from the Kirani family for generously supporting the seminar series. Hosts: Aditya Subramanian, Aditya Abhay Lonkar, Rahul Madhavan & Rameesh Paul DTSTART:20230109T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230110T120000Z UID:98b6c0e58b126c3134a5ca105b7ac1b7-388 DTSTAMP:19700101T120011Z DESCRIPTION:Hardness of Approximating Discrete Steiner Tree in L_p metrics URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/388/hardness-of-approximating-discrete-steiner-tree-in-l_p-metrics/ SUMMARY:In the Discrete Steiner Tree problem (DST), we are given as input two sets of points in a metric space, called terminals and facilities respectively, and the goal is to find the minimum-cost tree connecting the terminals, by possibly introducing new points (called Steiner points) from the set of facilities, as nodes in the solution. It was known that DST is APX hard in the L_1 metric by Trevisan (SICOMP 00) and in the Graph metric (and consequently, L_infinity metric ) by Chlebík and Chlebíková (TCS 08). It was open to rule out PTAS for DST in every other popular metric. In this talk, I will sketch the proof of APX hardness of DST in every Lp metric (in particular the Euclidean metric), edit metric, and Ulam metric. The talk is based on joint work with Henry Fleischmann and Surya Teja Gavva. Speaker Website https://karthikcs.org/ Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d Acknowledging the support from the Kirani family for generously supporting the seminar series. Aditya Subramanian, Aditya Abhay Lonkar, Rahul Madhavan & Rameesh Paul DTSTART:20230110T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230106T120000Z UID:a13dea7a1e8a1ff480617265b0464a9a-389 DTSTAMP:19700101T120014Z DESCRIPTION:Seismic Shifts: Challenges and Opportunities in the Post ISA Era of Computer Systems Design URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/389/seismic-shifts-challenges-and-opportunities-in-the-post-isa-era-of-computer-systems-design/ SUMMARY:For decades, Moores Law and its partner Dennard Scaling have together enabled exponential computer systems performance improvements at manageable power dissipation. With the slowing of Moore/Dennard improvements, designers have turned to a range of approaches for extending scaling of computer systems performance and power efficiency. These include specialized accelerators and heterogeneous parallelism. Unfortunately, the scaling gains afforded by these techniques come with significant costs: increased hardware and software complexity, degraded programmability and portability, and increased likelihood of design errors and security vulnerabilities. The long-held hardware-software abstraction offered by the Instruction Set Architecture (ISA) interface is fading quickly in this post-ISA era. The talk will cover a range of design opportunities and challenges, with a particular emphasis on my groups recent work on automated full-stack verification, security analysis, and the surprising alignments between full-stack issues in both classical and quantum computing systems. DTSTART:20230106T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230106T120000Z UID:05d47ae86961acad93e9522039098c48-390 DTSTAMP:19700101T120016Z DESCRIPTION:The Right to Deny URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/390/the-right-to-deny/ SUMMARY:Plausible deniability seems like the ultimate get-out-of-jail-free card. But how can we make it work when it comes to digital information sent in a public network. Deniable encryption, defined by Canetti et al (Crypto 1997), suggests a method to achieve deniability by the sender of encrypted messages to overcome this problem. The idea is especially interesting in the context of electronic elections to eliminate the threat of vote buying after a vote has been cast. I will present two new works on the subject. 1) With Agarwal and S. Mossel (Crypto21) we define and construct sender Deniable Fully Homomorphic Encryption with compact ciphertexts based on the Learning With Errors (LWE) polynomial hardness assumption. Deniable FHE enables storing encrypted data in the cloud to be processed securely without decryption, maintaining deniability of the encrypted data. 2) With Coladangelo and Vazirani (STOC22), we show a sender deniable encryption scheme where the encryption scheme is a quantum algorithm but the ciphertext is classical which is secure under the LWE polynomial hardness assumption. This scheme achieves for the first time simultaneously compactness, negligible deniability and polynomial encryption timeunder LWE. Furthermore, it is possible to extend the scheme so that coercion in an election cannot take place when the coercer is able to dictate all inputs to the deniable encryption algorithm even prior to encryption. DTSTART:20230106T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230106T120000Z UID:e3655dc2ad6591060e6fecec7e437f8c-391 DTSTAMP:19700101T120016Z DESCRIPTION:The Right to Deny URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/391/the-right-to-deny/ SUMMARY:Plausible deniability seems like the ultimate get-out-of-jail-free card. But how can we make it work when it comes to digital information sent in a public network. Deniable encryption, defined by Canetti et al (Crypto 1997), suggests a method to achieve deniability by the sender of encrypted messages to overcome this problem. The idea is especially interesting in the context of electronic elections to eliminate the threat of vote buying after a vote has been cast. I will present two new works on the subject. 1) With Agarwal and S. Mossel (Crypto21) we define and construct sender Deniable Fully Homomorphic Encryption with compact ciphertexts based on the Learning With Errors (LWE) polynomial hardness assumption. Deniable FHE enables storing encrypted data in the cloud to be processed securely without decryption, maintaining deniability of the encrypted data. 2) With Coladangelo and Vazirani (STOC22), we show a sender deniable encryption scheme where the encryption scheme is a quantum algorithm but the ciphertext is classical which is secure under the LWE polynomial hardness assumption. This scheme achieves for the first time simultaneously compactness, negligible deniability and polynomial encryption timeunder LWE. Furthermore, it is possible to extend the scheme so that coercion in an election cannot take place when the coercer is able to dictate all inputs to the deniable encryption algorithm even prior to encryption. DTSTART:20230106T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230113T120000Z UID:da3a3ebd780be6ba6a6256873244f07c-392 DTSTAMP:19700101T120011Z DESCRIPTION:Visualization - Data analysis with the human in the loop. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/392/visualization-data-analysis-with-the-human-in-the-loop/ SUMMARY:Effective analysis of increasingly large and complex data from simulations and experiments is a major step in the scientific process. If understanding or knowledge generation is the major goal of the process it is essential to keep the scientist in the loop. Practically this means building environments for scientific reasoning through interactive exploration of the data. This requires an effective interplay of automatic analysis methods providing some guidance through appropriate data abstraction and interaction methods that give the user control over the analysis process. Solutions must be found in close collaboration with the domain experts while developing generic tools and concepts that also can be adapted to other applications. In this talk, I will discuss a few visual analysis applications from our recent research including use cases from engineering, chemistry, and medicine. DTSTART:20230113T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230113T120000Z UID:6413cb3ee326b1884321fd535a4c8898-393 DTSTAMP:19700101T120016Z DESCRIPTION:Models of Simplicity: Encounters in power systems, power electronics, and physiology URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/393/models-of-simplicity-encounters-in-power-systems-power-electronics-and-physiology/ SUMMARY:Professor V.V.S. Sarma ("VVS") and his collaborators and students used, extended, and applied pattern recognition and machine learning/AI ideas and methods in a variety of areas, starting in the mid 1970's, but his interests expanded beyond: he wrote that "while my thoughts were on the philosophy of AI, my students were working on engineering aspects of AI and incorporation of AI ideas in system design." The VVS group's applications of AI in each domain - whether for speaker recognition, aircraft maintenance protocols, computer system reliability, hot forging processes in metallurgy, or remote sensing - were built on careful understanding of the application area (he refers to hours of "domain knowledge elicitation sessions"), to extract and exploit appropriate models for system design. <br> <br> I began my academic career with a grounding similar to VVS's, in electrical engineering, and subsequently dynamic systems, control theory, and signal processing. Power systems and power electronics were my focus for 20+ years, but a sabbatical at a Boston hospital marked a switch to "computational physiology" for clinical inference. I seem to have landed in particular applications in these domains where the most elementary models have been disproportionately insightful and effective (with only tangential help from machine learning), and I will talk about these. <br> <br> Biosketch: George Verghese earned his BTech from IITM in '74, his MS from Stony Brook University in '75, and his PhD (under Prof. Thomas Kailath) from Stanford in '79, all in electrical engineering. He has been with the EECS Department at MIT ever since, where he is a chaired professor of electrical and biomedical engineering, and has won treasured MIT-wide awards for undergraduate education and for mentoring. He is an IEEE Fellow, and coauthor of Signals, Systems and Inference (2015, with Oppenheim) and Principles of Power Electronics (2nd edition, with Kassakian, Perreault and Schlecht, publication mid-2023, a mere 32 years after the 1st edition). <br> George Verghese earned his BTech from IITM in 74, his MS from Stony Brook University in 75, and his PhD (under Prof. Thomas Kailath) from Stanford in 79, all in electrical engineering. He has been with the EECS Department at MIT ever since, where he is a chaired professor of electrical and biomedical engineering, and has won treasured MIT-wide awards for undergraduate education and for mentoring. He is an IEEE Fellow, and coauthor of Signals, Systems and Inference (2015, with Oppenheim) and Principles of Power Electronics (2nd edition, with Kassakian, Perreault and Schlecht, publication mid-2023, a mere 32 years after the 1st edition). DTSTART:20230113T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230113T120000Z UID:5659c614375d76d3440fe1e6e2643ba9-394 DTSTAMP:19700101T120017Z DESCRIPTION:Low Degree Testing over the Reals URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/394/low-degree-testing-over-the-reals/ SUMMARY:We study the problem of testing whether a function $f: reals^n to reals$ is a polynomial of degree at most $d$ in the emph{distribution-free} testing model. Here, the distance between functions is measured with respect to an unknown distribution $mathcal{D}$ over $reals^n$ from which we can draw samples. In contrast to previous work, we do not assume that $mathcal{D}$ has finite support. We design a tester that given query access to $f$, and sample access to $mathcal{D}$, makes $poly(d/eps)$ many queries to $f$, accepts with probability $1$ if $f$ is a polynomial of degree $d$, and rejects with probability at least $mathfrac{2}{3}$ if every degree-$d$ polynomial $P$ disagrees with $f$ on a set of mass at least $eps$ with respect to $mathcal{D}$. Our result also holds under mild assumptions when we receive only a polynomial number of bits of precision for each query to $f$, or when $f$ can only be queried on rational points representable using a logarithmic number of bits. Along the way, we prove a new stability theorem for multivariate polynomials that may be of independent interest. This is a joint work with Arnab Bhattacharyya, Esty Kelman, Noah Fleming, and Yuichi Yoshida, and will appear in SODA23. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Aditya Subramanian, Aditya Abhay Lonkar, Rahul Madhavan & Rameesh Paul DTSTART:20230113T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230313T120000Z UID:3cfbe0ffbf6924cd07dc30ae0c58dc2a-395 DTSTAMP:19700101T120011Z DESCRIPTION:A Case for Correctly Rounded Math Libraries URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/395/a-case-for-correctly-rounded-math-libraries/ SUMMARY:This talk will provide an overview of the RLIBM project where we are building a collection of correctly rounded elementary functions for multiple representations and rounding modes. Historically, polynomial approximations for elementary functions have been designed by approximating the real value. In contrast, we make a case for approximating the correctly rounded result of an elementary function rather than the real value of an elementary function in the RLIBM project. Once we approximate the correctly rounded result, there is an interval of real values around the correctly rounded result such that producing a real value in this interval rounds to the correct result. This interval is the freedom that the polynomial approximation has for an input, which is larger than the ones with the mini-max approach. Using these intervals, we structure the problem of generating polynomial approximations that produce correctly rounded results for all inputs as a linear programming problem. The results from the RLIBM project makes a strong case for mandating correctly rounded results at least for any representation that has fewer than or equal to 32-bits. DTSTART:20230313T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230118T120000Z UID:27dd3e8421d70537fee43112a1ec8c25-396 DTSTAMP:19700101T120011Z DESCRIPTION:Average Reward Actor-Critic with Deterministic Policy Search URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/396/average-reward-actor-critic-with-deterministic-policy-search/ SUMMARY:The average reward criterion is relatively less explored as most existing works in the Reinforcement Learning literature consider the discounted reward criterion. There are few recent works that present on-policy average reward actor-critic algorithms, but average reward off-policy actor-critic is relatively less explored. In this work, we present both on-policy and off-policy deterministic policy gradient theorems for the average reward performance criterion. Using these theorems, we also present an Average Reward Off-Policy Deep Deterministic Policy Gradient (ARO-DDPG) Algorithm. We first show asymptotic convergence analysis using the ODE-based method. Subsequently, we provide a finite time analysis of the resulting stochastic approximation scheme with linear function approximator and obtain an $epsilon$-optimal stationary policy with a sample complexity of $Omega(epsilon^{-2.5})$. We compare the average reward performance of our proposed ARO-DDPG algorithm and observe better empirical performance compared to state-of-the-art on-policy average reward actor-critic algorithms over MuJoCo-based environments. DTSTART:20230118T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230131T120000Z UID:02bf6d39f69776dc8d7ab1e1c3c7786f-397 DTSTAMP:19700101T120011Z DESCRIPTION:Improved Approximation Bounds On Maximum Edge q-coloring Of Dense Graphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/397/improved-approximation-bounds-on-maximum-edge-q-coloring-of-dense-graphs/ SUMMARY:The {it anti-Ramsey number} $ar(G,H)$ with {it input graph} $G$ and {it pattern graph} $H$, is the maximum positive integer $k$ such that there exists an edge coloring of $G$ using $k$ colors, in which there are no {it rainbow} subgraphs isomorphic to $H$ in $G$. ($H$ is {it rainbow} if all its edges get distinct colors). The concept of {it anti-Ramsey number} was introduced by Erd os, Simanovitz, and S'os in 1973. Thereafter several researchers investigated this concept in the {it combinatorial} setting. The cases where pattern graph $H$ is a complete graph $K_r$, a path $P_r$ or a star $K_{1,r}$ for a fixed positive integer $r$, are well studied. Recently, Feng et al. revisited the {it anti-Ramsey} problem for the pattern graph $K_{1,t}$ (for $t ge 3$) purely from an {it algorithmic} point of view, due to its applications in {it interference modeling} of {it wireless networks}. They posed it as an optimization problem, the emph {maximum edge $q$-coloring problem.} For a graph $G$ and an integer $qgeq 2$, an edge $q$-coloring of $G$ is an assignment of colors to edges of $G$, such that edges incident on a vertex span at most $q$ distinct colors. The {it maximum edge} $q$-coloring problem seeks to {it maximize} the number of colors in an edge $q$-coloring of the graph $G$. Note that the {it optimum value} of the edge $q$-coloring problem of $G$ equals $ar(G,K_{1,q+1})$. We study $ar(G,K_{1,t})$, the {it anti-Ramsey number} of stars, for each fixed integer $tgeq 3$, both from {it combinatorial} and {it algorithmic} point of view. The first of our main results, presents an upper bound for $ar(G,K_{1,q+1})$, in terms of number of vertices and the minimum degree of $G$. The second one improves this result for the case of triangle free input graphs. For a positive integer $t$, let $H_t$ denote a subgraph of $G$ with maximum number of possible edges and maximum degree $t$. From an observation of Erd"os, Simanovitz, and S'os, we get: $|E(H_{q-1})| + 1leq ar(G,K_{1,q+1}) leq |E(H_{q})|$. For instance, when $q=2$, the subgraph $E(H_{q-1})$ refers to a maximum matching. It looks like $|E(H_{q-1})|$ is the most natural parameter associated with the anti-ramsey number $ar(G,K_{1,q+1})$ and the approximation algorithms for the maximum edge coloring problem proceed usually by first computing the $H_{q-1}$, then coloring all its edges with different colors and by giving one (sometimes more than one) extra colors to the remaining edges. The approximation guarantee of these algorithms usually depend on upper bounds for $ar(G,K_{1,q+1})$ in terms of $|E(H_{q-1})|$. Our third main result presents an upper bound for $ar(G,K_{1,q+1})$ in terms of $|E(H_{q-1})|$. All our results have algorithmic consequences. For some large special classes of graphs, such as $d$-regular graphs, where $dgeq 4$, our results can be used to prove a better approximation guarantee for the sub-factor based algorithm (Algorithm~ref{alg:feng}). We also show that all our bounds are almost tight. Results for the case $q=2$ were done earlier by Chandran et al cite{chandran18}. In this thesis, we extend it further for each fixed integer $q > 2$. DTSTART:20230131T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230116T120000Z UID:081548122d10eefcccf82009f088f622-398 DTSTAMP:19700101T120016Z DESCRIPTION:Automated Decision Making for Safety Critical Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/398/automated-decision-making-for-safety-critical-applications/ SUMMARY:Building robust decision making systems for autonomous systems is challenging. Decisions must be made based on imperfect information about the environment and with uncertainty about how the environment will evolve. In addition, these systems must carefully balance safety with other considerations, such as operational efficiency. Typically, the space of edge cases is vast, placing a large burden on human designers to anticipate problem scenarios and develop ways to resolve them. This talk discusses major challenges associated with ensuring computational tractability and establishing trust that our systems will behave correctly when deployed in the real world. We will outline some methodologies for addressing these challenges and point to some research applications that can serve as inspiration for building safer systems. DTSTART:20230116T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230120T120000Z UID:6feb5d28d3cce91ead7afae27c231155-400 DTSTAMP:19700101T120016Z DESCRIPTION:Energy-efficient 2.5D Architectures with Processing-in-memory for Machine Learning Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/400/energy-efficient-2-5d-architectures-with-processing-in-memory-for-machine-learning-applications/ SUMMARY:Processing-in-memory (PIM) is a promising technique to accelerate deep learning (DL) workloads. Emerging DL workloads (e.g., ResNet with 152 layers) consist of millions of parameters, which increase the area and fabrication cost of monolithic PIM accelerators. The fabrication cost challenge can be addressed by 2.5-D systems integrating multiple PIM chiplets connected through a network-on-package (NoP). However, server-scale scenarios simultaneously execute multiple compute-heavy DL workloads, leading to significant inter-chiplet data volume. State-of-the-art NoP architectures proposed in the literature do not consider the nature of DL workloads. In this talk, we will discuss a novel server-scale 2.5-D manycore architecture that accounts for the traffic characteristics of DL applications. Comprehensive experimental evaluations with different system sizes as well as diverse emerging DL workloads demonstrate that the architecture achieves significant performance and energy consumption improvements with much lower fabrication cost than state-of-the-art NoP topologies. DTSTART:20230120T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230210T120000Z UID:43528966f5417c7f4bef358437e85ca7-401 DTSTAMP:19700101T120011Z DESCRIPTION:IASO: A Fail-Slow Detection and Mitigation Framework for Distributed Storage Services URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/401/iaso-a-fail-slow-detection-and-mitigation-framework-for-distributed-storage-services/ SUMMARY:Distributed systems, whether large or small, have to handle two kinds of failures. The first type is called "fail-stop" failure where a component, system, or process might stop operating completely. These are easy to detect and have fairly standard ways of being handled. The second type of failure is called "fail-slow" failure and is characterized by some part of the system experiencing degraded performance, but is not completely non-functional. Such failures are much harder to detect and can cause a series of cascading failures. The talk focuses on the detection, mitigation and resolution of such fail-slow failures through a framework conceived and built in Nutanix, called IASO. IASO is a peer-based, non-intrusive fail-slow detection framework that Nutanix has deployed in customer sites and which has helped mitigate a large number of incidents before they cascaded into complete cluster outages. The talk focuses on the design of IASO with highlight on how the various choices made are essential in a real-world setting. This talk is based on a paper that presented at Usenix ATC 2019. DTSTART:20230210T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230127T120000Z UID:06299168c9cc99095dddb696c3e593c5-402 DTSTAMP:19700101T120014Z DESCRIPTION:Visualization research - from data analysis to science communication URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/402/visualization-research-from-data-analysis-to-science-communication/ SUMMARY:Visualization is omnipresent in everyday life serving many different purposes, examples range from plots in the newspapers to illustrations in textbooks. However, visualization goes far beyond such examples and pretty pictures. Visual data analysis has developed into an essential component of modern scientific workflows supporting understanding and reasoning about data. In this talk, I will mainly focus on the use of visualization for data analysis and exploration, and conclude with an outlook on how similar methods can be used in science communication. <br> As data is increasingly large and complex, effective data exploration requires abstractions that serve as a backbone for easy navigation through data. To this end, topological data analysis (TDA) has proven to provide fundamental tools in visualzaiton applicaitons. It provides multi-scale data summaries with nice mathematical properties and guarantees. In the talk, I will demonstrate a few examples using topological descriptors for feature tracking in time-dependent scalar fields. The examples include cyclone evolutions in weather modeling and structure tracking in flow simulations. DTSTART:20230127T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230202T120000Z UID:70dc506294cbc8ce335a7667551c6db1-403 DTSTAMP:19700101T120018Z DESCRIPTION:On Computing Homological Hitting Sets URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/403/on-computing-homological-hitting-sets/ SUMMARY:Cut problems form one of the most fundamental classes of problems in algorithmic graph theory. In this paper, we initiate the algorithmic study of a high-dimensional cut problem. The problem we study, namely, Homological Hitting Set (HHS), is defined as follows: Given a nontrivial r-cycle z in a simplicial complex, find a set S of r-dimensional simplices of minimum cardinality so that S meets every cycle homologous to z. Our first result is that HHS admits a polynomial-time solution on triangulations of closed surfaces. Interestingly, the minimal solution is given in terms of the cocycles of the surface. Next, we provide an example of a 2-complex for which the (unique) minimal hitting set is not a cocycle. Furthermore, for general complexes, we show that HHS is W[1]-hard with respect to the solution size p. In contrast, on the positive side, we show that HHS admits an FPT algorithm with respect to p + d, where d is the maximum degree of the Hasse graph of the complex K. <br> The talk is based on joint work with Ulrich Bauer and Meirav Zehavi. <br> <br> Microsoft teams link: <br> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d <br> <br> We acknowledge the Kirani familys generous support towards conducting this seminar series. <br> <br> Hosts: Aditya Abhay Lonkar, Aditya Subramanian, Rahul Madhavan & Rameesh Paul DTSTART:20230202T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230206T120000Z UID:0a43eeb68066e7fff3ed4d18fdfc1fcf-404 DTSTAMP:19700101T120014Z DESCRIPTION:Exploring the Size of Order-k Voronoi Tessellation and Chromatic Delaunay Mosaic URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/404/exploring-the-size-of-order-k-voronoi-tessellation-and-chromatic-delaunay-mosaic/ SUMMARY:In this presentation, we will first delve into the topic of the size of order-k Voronoi tessellations. Specifically, we will examine how Lee's inductive argument for counting cells in R^2 can be generalized to R^3, resulting in precise relations involving Morse-theoretic quantities for piecewise constant functions on planar arrangements. Additionally, we will introduce the concept of a chromatic Delaunay mosaic, which is a Delaunay mosaic in R^(s+d) that illustrates the intermixing of points of (s+1) colors within a locally finite set of points in R^d. Our primary findings include bounds on the size of this chromatic Delaunay mosaic. These results are the product of a collaborative effort with Sebastiano Cultrera, Herbert Edelsbrunner, Ondrej Draganov, and Morteza Saghafian. <br> <br> This is an online Seminar. The Teams URL for this is: <br> <a href="https://tinyurl.com/RanitaBiswasTalk">https://tinyurl.com/RanitaBiswasTalk</a> DTSTART:20230206T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230201T120000Z UID:e532813b563acb4b01e2e164143edbd6-405 DTSTAMP:19700101T120011Z DESCRIPTION:CodeQueries: Benchmarking Query Answering over Source Code URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/405/codequeries-benchmarking-query-answering-over-source-code/ SUMMARY:Software developers often make queries about the security, performance effectiveness, and maintainability of their code. Through an iterative debugging process, developers analyze the code to find answers to these queries. The process can be seen as a question-answering task that requires developers to identify code spans satisfying certain properties. Many of these queries can be answered by existing code analysis tools such as CodeQL. However, using such tools requires design, implementation, and verification efforts. <br> In this work, we propose an alternative to the code analysis tools by formulating the task of query answering over source code as a span prediction problem. In the proposed approach, a neural model is designed to predict appropriate answer spans in a code in response to a query. The required supporting-facts to justify the predicted answers are also identified by the model. Pre-trained language models for code are fine-tuned on a newly prepared challenging dataset, CodeQueries, for query answering over source code. We demonstrate that the proposed approach performs well on the query answering over source code task when only relevant code blocks are provided as input to the model. Experiments conducted on the dataset demonstrate that the proposed neural approach is robust to noisy span labeling and can even handle code with minor syntax errors. Although large-sized code and limited training examples adversely affect the model performance, we suggest methods to address these issues. Based on our study, we believe that the proposed neural approach will be an additional tool in a developers toolbox for query answering over source code. DTSTART:20230201T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230203T120000Z UID:784b81be76f96d33024a1980fa94380f-406 DTSTAMP:19700101T120011Z DESCRIPTION:µIRs-Intermediate Representation for Agile Design of Accelerators URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/406/a%c2%b5irs-intermediate-representation-for-agile-design-of-accelerators/ SUMMARY:Creating high quality application-specific accelerators requires us to make iterative changes to both algorithm behavior and microarchitecture, and this is a tedious and error-prone process. We propose a generalized intermediate representation for describing accelerator microarchitecture, μIR, and an associated pass framework, μopt. μIR represents the accelerator as a concurrent structural graph in which the components roughly correspond to microarchitecture level hardware blocks (e.g., function units, network, memory banks). There are two important benefits i) it decouples microarchitecture optimizations from algorithm/program optimizations. ii) it decouples microarchitecture optimizations from the RTL generation. Computer architects express their ideas as a set of iterative transformations of the μIR graph that successively refine the accelerator architecture. The μIR graph is then translated to Chisel, while maintaining the execution model and cycle-level performance characteristics. We study three broad classes of optimizations: Timing (e.g., Pipeline re-timing), Spatial (e.g., Compute tiling), and Higher-order Ops (e.g., Tensor function units) that deliver between 1.5 — 8× improvement in performance; overall 5—20× speedup compared to an ARM A9 1Ghz. DTSTART:20230203T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230209T120000Z UID:61bc4f2cab82b5fd4a23e6d8b796f2c0-407 DTSTAMP:19700101T120016Z DESCRIPTION:Algorithms for Achieving Fairness and Efficiency in Matching Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/407/algorithms-for-achieving-fairness-and-efficiency-in-matching-problems/ SUMMARY:Matching problems arise in numerous practical settings. Fairness and efficiency are two desirable properties in most such real-world settings. This dissertation work presents new approaches and algorithms to identify fair and/or efficient matchings. The thesis is organised into two logical parts: two sided preferences and single sided preferences. <br> Part 1: Two Sided Preferences <br> Incentive Compatibility in Stable Fractional Matchings <br> We investigate the existence of incentive compatible mechanisms that find stable fractional matchings. We show, for general settings, that no incentive compatible mechanism can be stable. We characterise the space of instances that have a unique stable fractional matching. We prove for this set of instances that any stable matching mechanism will be incentive compatible. <br> Fairness and Stability in Many-to-One Matchings <br> We seek to optimize a fairness measure over the space of stable many-to-one matchings, motivated by a college admissions setting. With leximin optimality as the fairness notion, we first show the intractability of this problem. We identify a reasonable set of assumptions that makes this problem solvable in polynomial time. This requires that the agents on either side have the same ordinal rankings over the agents on the other side and that these preferences are strict. We show that on relaxing to weak rankings, the problem becomes APX-Hard. When we remove the ranking assumption but maintain strict preferences, the problem is NP-Hard. We show that the leximin optimal stable matching can be efficiently computed in the special case of two colleges. <br> Part 2: Single Sided Preferences <br> Repeated Matchings <br> We propose a novel repeated matching model where the valuations of agents may change with how often they have received an item in the past. We study achieving fairness and efficiency separately as well as jointly in this setting. We find that optimizing for social welfare is NP-Hard for general valuations and tractable when the valuations are monotone with time. We also prove that maximizing for social welfare over the space of EF1 repeated matchings is NP-Hard. Further, we provide algorithms and non-existence results for EF1 and EFX repeated matchings in different settings. <br> Fair and Efficient Delivery <br> Motivated by the classical delivery problem, we introduce a novel model of fair division where delivery tasks must be fairly distributed among a set of agents. The delivery tasks are placed on the vertices of a given acyclic graph. The cost incurred by the agents is determined by the distance they travel from the hub where they start to service their assigned tasks. We study the existence of fair and efficient allocations of tasks to agents. We choose the fairness notions: EF1 and MMS, and efficiency notions: Pareto optimality and Social optimality. We find that while all these notions can be satisfied independently, the only combination of fairness and efficiency that can always be guaranteed is MMS and PO. For the remaining combinations, we provide characterisations of the space of instances for which they can be achieved. We find that most of the relevant problems are NP-Hard. We provide an XP-algorithm which finds the different combinations of fairness and efficiency whenever they exist. DTSTART:20230209T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230221T120000Z UID:6a55797c72b106054fc27473d5fcaa78-409 DTSTAMP:19700101T120016Z DESCRIPTION:Time-SpaceTradeoffs for Collisions in Hash Functions URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/409/time-spacetradeoffs-for-collisions-in-hash-functions/ SUMMARY:Cryptographic hash functions are functions that take arbitrary length inputs and output fixed length digest. They are one of the most important cryptographic primitives and widely used in applications today. Apart from the compression requirement, the applications using these functions could need additional properties to be provably secure. One such, perhaps the most important property is collision resistance. This property has been well studied for uniform adversaries. However, uniform adversaries fail to capture many real-world adversaries. Hence, several recent works have studied the collision resistance property for non-uniform adversaries. Analyzing non-uniform adversaries presents several challenges. That is why Dodis et al in their EUROCRYPT 18 paper presented a reduction to another (easier to analyze) model named Bit-fixing model. In our CRYPTO 20 paper, we showed that adversaries in this Bit-fixing model are too strong when the length of the collisions are bounded. We also showed a reduction to the Multi-instance model, which helped us obtain better results for restricted parameter ranges. In our recent CRYPTO 22 paper, we further explored the relation between the Bit-fixing model and the Multi-instance model and further improved the results with our new findings. The talk will include some preliminary definitions, detailed description of these models, results and a high level idea of the techniques from all the relevant works. DTSTART:20230221T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230215T120000Z UID:872f25447b232db4851e10358a648cae-410 DTSTAMP:19700101T120011Z DESCRIPTION:Privadome: A System for Citizen Privacy in the Delivery Drone Era URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/410/privadome-a-system-for-citizen-privacy-in-the-delivery-drone-era/ SUMMARY:E-commerce companies are actively considering the use of delivery drones for customer fulfillment, leading to growing concerns around citizen privacy. Drones are equipped with cameras, and the video feed from these cameras is often required as part of routine navigation, be it for semi-autonomous or fully-autonomous drones. Footage of ground based citizens captured in these videos may lead to privacy concerns. <br> This M.Tech. (Research) thesis presents the design, implementation and evaluation of Privadome, a system that implements the vision of a virtual privacy dome centered around the citizen. Privadome is designed to be integrated with city-scale regulatory authorities that oversee delivery drone operations and realizes this vision through two components, Pd-Mpc and Pd-Ros. Pd-Mpc allows citizens equipped with a mobile device to identify drones that have captured their footage. It uses secure two-party computation to achieve this goal without compromising the privacy of the citizens location. Pd-Ros allows the citizen to communicate with such drones and obtain an audit trail showing how the drone uses their footage and determine if privacy- preserving steps are taken to sanitize the footage. An experimental evaluation of Privadome shows that the system scales to near-term city- scale delivery drone deployments (hundreds of drones). We show that with Pd-Mpc the mobile data usage on the citizens mobile device is comparable to that of routine activities on the device, such as streaming videos. We also show that the workflow of Pd-Ros consumes a modest amount of additional CPU resources and power on our experimental platform. DTSTART:20230215T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230210T120000Z UID:402a91adc7b424068431b8154e34f0f6-411 DTSTAMP:19700101T120011Z DESCRIPTION:Equivalence Test for Read-Once Arithmetic Formulas URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/411/equivalence-test-for-read-once-arithmetic-formulas/ SUMMARY:We study the polynomial equivalence problem for orbits of read-once arithmetic formulas (ROFs). Read-once formulas have received considerable attention in both algebraic and Boolean complexity and have served as a testbed for developing effective tools and techniques for analyzing circuits. Two n-variate polynomials f,g∈F[x] are equivalent, denoted as f∼g, if there is an A∈GL(n,F) such that f=g(Ax). The orbit of f is the set of all polynomials equivalent to f. We investigate the complexity of the following two natural problems on ROFs: 1. Equivalence test for ROFs: Given black-box access to f, check if it is in the orbit of an ROF. If yes, output an ROF C and an A∈GL(n,F) such that f=C(Ax). 2. Polynomial equivalence for orbits of ROFs: Given black-box access to f and g in the orbits of two unknown ROFs, check if f∼g. If yes, output an A∈GL(n,F) such that f=g(Ax). These problems are significant generalizations of two well-studied problems in algebraic complexity, namely reconstruction of ROFs and quadratic form equivalence. In this work, we give the first randomized polynomial-time algorithms (with oracle access to quadratic form equivalence) to solve the two problems. The equivalence test works for general ROFs; it also implies an efficient learning algorithm for random arithmetic formulas of unbounded depth and fan-in (in the high number of variables setting). The algorithm for the second problem, which invokes the equivalence test, works for mildly restricted ROFs, namely additive-constant-free ROFs. The equivalence test is based on a novel interplay between the factors and the essential variables of the Hessian determinant of an ROF, the essential variables of the ROF, and certain special structures in the ROF that we call DTSTART:20230210T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230214T120000Z UID:ced05f90bcf82077d626d66af1578153-412 DTSTAMP:19700101T120011Z DESCRIPTION:Proving and Programming with PVS URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/412/proving-and-programming-with-pvs/ SUMMARY:SRI's Prototype Verification System is an interactive proof assistant that has been in active development and use for over thirty years. PVS features a specification language based on a richly typed higher-order logic with algebraic datatypes, dependent predicate subtypes, and parametric theories. The interactive theorem prover employs a range of automated proof strategies for simplification, rewriting, and case analysis, along with built-in decision procedures for SAT and SMT solving. The applicative fragment of PVS can be viewed as a functional programming language, and executable code can be generated in Common Lisp and C, among other languages. PVS includes extensive libraries spanning a range of topics in mathematics and computing. Since the formalizations include a significant amount of computational content in the form of executable programs, it is useful to generate code from the programs that executes efficiently. The generated code can be integrated as verified components within larger systems or employed as reference implementations. The talk is an informal overview of the underlying theoretical foundations, and the proof and code generation capabilities of PVS. DTSTART:20230214T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230216T120000Z UID:5a82f8cb5814569c31f88c4746cf29cc-413 DTSTAMP:19700101T120018Z DESCRIPTION:Machine Learning and Logic: Fast and Slow Thinking URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/413/machine-learning-and-logic-fast-and-slow-thinking/ SUMMARY:Computer science seems to be undergoing a paradigm shift. Much of earlier research was conducted in the framework of well-understood formal models. In contrast, some of the hottest trends today shun formal models and rely on massive data sets and machine learning. A cannonical example of this change is the shift in AI from logic programming to deep learning. I will argue that the correct metaphore for this development is not paradigm shift, but paradigm expansion. Just as General Relativity augments Newtonian Mechanics, rather than replace it -- we went to the moon, after all, using Newtonian Mechanics -- data-driven computing augments model-driven computing. In the context of Artificial Intelligence, machine learning and logic correspond to the two modes of human thinking: fast thinking and slow thinking. The challenge today is to integrate the model-driven and data-driven paradigms. I will describe one approach to such an integration -- making logic more quantitative. <br> <br> For more details please visit: https://www.csa.iisc.ac.in/theoryseminars/?talk=20230216_MosheVardi <br> Link to online talk: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d DTSTART:20230216T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230301T120000Z UID:c51a5425c61abf19f93426979f15ebe2-414 DTSTAMP:19700101T120011Z DESCRIPTION:Understanding Performance of Internet Video using Network Measurement Data URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/414/understanding-performance-of-internet-video-using-network-measurement-data/ SUMMARY:Internet video is used extensively for entertainment, education, work, and telehealth, making it the largest contributor to Internet traffic. Due to these two factors, i.e., popularity and resource utilization, it is crucial to understand the network dynamics of Internet video. Understanding network behavior of the video, however, has several challenges, including (1). complex interactions among entities on the network path (e.g., CDNs, ISPs) and within the vertical network stack, (2). limited information available to some stakeholders (e.g., network operators), and (3). heterogeneity in network and user contexts. In this talk, I will cover two of my research directions that use network data to provide insights into the dynamic behavior of the Internet video applications. The first is on using passive measurements to enable network operators to understand Quality of Experience metrics for HTTP-based adaptive streaming (e.g., Netflix, YouTube). The second is on using empirical network data to understand performance and network utilization of video conferencing applications (e.g., Zoom). The talk will conclude with a brief discussion of my ongoing work on mapping broadband inequity and future research plans. DTSTART:20230301T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230217T120000Z UID:ee9a203348b3db7386298a347b9c77f4-415 DTSTAMP:19700101T120011Z DESCRIPTION:Towards Robustness of Neural Legal Judgment System URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/415/towards-robustness-of-neural-legal-judgment-system/ SUMMARY:Legal Judgment Prediction (LJP) implements Natural Language Processing (NLP) techniques to predict judgment results based on fact description. It can play a vital role as a legal assistant and benefits legal practitioners and regular citizens. Recently, the rapid advances of transformer-based pretrained language models led to considerable improvement in this area. However, empirical results show that existing LJP systems are not robust to adversaries and noise. Also, they cannot handle large-length legal documents. In this work, we explore the robustness and efficiency of LJP systems even in a low data regime. <br> In the first part, we empirically verified that existing state-of-the-art LJP systems are not robust. We further provide our novel architecture for LJP tasks which can handle extensive text lengths and adversarial examples. Our model performs better than state-of-the-art models, even in the presence of adversarial examples of the legal domain.<br> In the second part, we investigate the approach for the LJP system in a low data regime. We provide a novel architecture using a few-shot approach that is also robust to adversaries. We conducted extensive experiments on American, European, and Indian legal datasets in the few-shot scenario. Our model, though trained using the few-shot approach, performs as well as state-of-the-art models which are trained using large datasets in the legal domain. DTSTART:20230217T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230217T120000Z UID:3f2158b4ddad5e494810993bfcc87c5d-416 DTSTAMP:19700101T120012Z DESCRIPTION:Fragile Interpretations and Intepretable Models in NLP URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/416/fragile-interpretations-and-intepretable-models-in-nlp/ SUMMARY:Deployment of deep learning models in critical areas is still an issue of concern, as the cost of making a wrong decision is very high in these areas. As a result, the final decision in these settings is human-centric. Also, these models act as black boxes, and we are unaware of their internal workings. Therefore the models must be explainable to know their internal workings. So, if the model is explainable, is it safe to deploy it in real-world settings involving huge risks? <br> Our work centers around the concept of fragile interpretations considering the models robustness and the robustness of interpretations. We have proposed an algorithm that perturbs the input text and generates adversarial examples with the same prediction as the input but with different interpretations. Through our experiments, we have provided a detailed analysis of whether these interpretations are reliable and whether to trust the model or the interpretations. We have provided the reason for the fragility in the case of NLP. Taking this into account, we have proposed two interpretable models, one for a multi-task offensive language detection task and the other for a sentence-pair similarity detection task. DTSTART:20230217T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230224T120000Z UID:2ffcc372e420fb355f5d1cfe0bbe5388-418 DTSTAMP:19700101T120011Z DESCRIPTION:2D Expansion in Random Geometric Graphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/418/2d-expansion-in-random-geometric-graphs/ SUMMARY:High-dimensional expansion is a generalization of expansion to hypergraphs where the expansion of certain random walks are witnessed by local neighborhoods. <br> This talk is on the topic of constructing natural probability distributions over sparse high-dimensional expanders (HDXes). On one hand, (standard/1-dimensional) sparse expanders are plentiful, since a constant degree random graph is an expander with high probability. On the other hand, most natural random models over hypergraphs fail to exhibit 2-dimensional expansion unless the average degree exceeds sqrt(number of vertices). <br> However, sparse HDXes are known to exist due to algebraic and number theory based construction. We describe some progress towards constructing sparse HDXes based on random geometrics graphs on the unit sphere. <br> The talk is based on joint work with Siqi Liu (Berkeley), Tselil Schramm (Stanford), and Elizabeth Yang (Berkeley). <br> <br> Microsoft teams link: <br> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d <br> <br> We are grateful to the Kirani family for generously supporting the theory seminar series <br> <br> Hosts: Rameesh Paul, Aditya Subramanian, Aditya Abhay Lonkar and Rahul Madhavan DTSTART:20230224T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230228T120000Z UID:67abd79a602376cb3c5a5ea4d2dd29f6-419 DTSTAMP:19700101T120015Z DESCRIPTION:What is common to robots, proteins, genomics and video games? URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/419/what-is-common-to-robots-proteins-genomics-and-video-games/ SUMMARY:The once familiar story of machine learning, specifically deep-learning, facilitating inordinate progress widely across disciplines quickly evolved into the realisation that they are data hungry and difficult to explain or analyse. In my research, I explore the benefits of incorporating a physics model within the learner to alleviate some of these problems. Although this could be included under the popular buzz-phrase Physics Inspired AI, my approach has been to use rapid (hence approximate) physics models by learning the discrepancy between their approximations and an accurate (hence computationally intensive) simulation model. <br> The Computer Graphics community is curiously comfortable with the dichotomy between accurate versus timely algorithms that solve the same computational problems (usually physical simulation) under different constraints. In my research, I explore both of these strands, each with a different goal. On the one hand, I strive to develop formalisms with the goal of assessing inaccuracies of existing models. On the other, I investigate applications of `quick and dirty approximations for applications that impose a strict time-budget. In this talk, I will provide an overview of these goals and the tension between them. After an introduction to Edinburgh and the School of Informatics as an exciting venue for visiting students, I will present accuracy and timeliness as contrasting notions of error and their relative importance across applications. I will describe my general research directions using examples including a recent (SIGGRAPH 22) paper on the analysis of error in light transport and previous works on the use of approximate physics simulation for robotic manipulation. <br> In this talk I will be sharing problems, that I am excited by, in protein design and genomics and some progress that we have made. I will touch upon recent work from my group on approximate learning of an NP-hard problem [2], our method for protein design [3] that is in the top-three methods available and a recent paper that uses functional analysis to explain why fixed sampling methods (such as NeRFs) will hit a fundamental roadblock when used to approximate light transport [1]. If time permits, I will also present some of our work on spectral coarsening of simplicial complexes [4]. <br> I will end my talk with some insights on potential routes for Indian researchers who are interested in academic jobs in the UK. This is particularly relevant if you are finishing your PhDs and are looking to apply for a post-doctoral research position. <br> [1] https://homepages.inf.ed.ac.uk/ksubr/research.html#SIGG22<br> [2] https://homepages.inf.ed.ac.uk/ksubr/research.html#AAAI22<br> [3] https://arxiv.org/pdf/2109.07925.pdf<br> [4] https://arxiv.org/abs/2207.01146<br> DTSTART:20230228T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230306T120000Z UID:b512bbfbd1811f731407c87670f7f864-420 DTSTAMP:19700101T120011Z DESCRIPTION:Fusing AI and Formal Methods for Automated Synthesis URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/420/fusing-ai-and-formal-methods-for-automated-synthesis/ SUMMARY:We entrust large parts of our daily lives to computer systems, which are becoming increasingly more complex. Developing scalable yet trustworthy techniques for designing and verifying such systems is an important problem. In this talk, our focus will be on automated synthesis, a technique that uses formal specifications to automatically generate systems (such as functions, programs, or circuits) that provably satisfy the requirements of the specification. I will introduce a state-of-the-art functional synthesis algorithm that leverages artificial intelligence to provide an initial guess for the system and then uses formal methods to repair and verify the guess to synthesize a system that is correct by construction. I will conclude by exploring the potential for combining AI and formal methods to address real-world scenarios. DTSTART:20230306T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230303T120000Z UID:f7dfee40642e40b3a47b495e25a4ed19-422 DTSTAMP:19700101T120011Z DESCRIPTION:A decentralised algorithm for minimizing multi-agent congestion cost on a network URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/422/a-decentralised-algorithm-for-minimizing-multi-agent-congestion-cost-on-a-network/ SUMMARY:Consider a model wherein a given set of agents need to reach the goal node of a network. The cost for each agent on any link depends on the congestion on that link as well as on a cost component that is private to that agent. We propose a multi-agent congestion cost minimization (MACCM) algorithm for minimizing the total cost incurred by the agents. Our algorithm is fully decentralised, uses linear function approximations that addresses privacy of agents costs as well as scalability aspects and achieves sub-linear regret. Each agent maintains an estimate of the global objective function and the algorithm relies on a multi-agent version of extended value-iteration. We illustrate computations on a hard instance. Our model is a generalisation of a classical learning problem, the stochastic shortest path problem. This is a joint work with Prashant Trivedi. DTSTART:20230303T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230304T120000Z UID:a62ae8d3afe7be3b2fbda702bb4ce8c3-424 DTSTAMP:19700101T120010Z DESCRIPTION:How do recommendation systems work? And, what are their privacy implications? URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/424/how-do-recommendation-systems-work-and-what-are-their-privacy-implications/ SUMMARY:Have you ever wondered how good Spotify, Netflix, Amazon, .. give you such great recommendations for what to listen, watch, purchase, and live our lives? The magic behind their recommendations are the deep learning machine learning models. These models capture seemingly end-less amounts of information about our online behavior and transform these behaviors into embeddings for future recommendations. These machine learning models are massive (think of terabytes of data), trained on equally imposing set of training samples, called sparse features. Every click, purchase, and even a mouse hover on a website is a sparse feature for training the model. In this talk I will first provide an overview of how current generation recommendation systems work. If these models can recommend so well, then, they must also know a lot about us. In fact, they do. By simply observing features such as click history and object interactions an attacker can de-anonymize users with extremely high probability, or track users across different interaction sessions. For instance, if you tell me what are the last two items you purchased online, I can track you with 97% accuracy. All of which is to say there is a lot of interesting privacy research that needs to get done. DTSTART:20230304T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230304T120000Z UID:81f3a4774f0e365d06e1615070dfba90-425 DTSTAMP:19700101T120015Z DESCRIPTION:Towards Sustainable Agriculture prioritizing Global South using Machine Learning URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/425/towards-sustainable-agriculture-prioritizing-global-south-using-machine-learning/ SUMMARY:The agricultural sector, which is a major source of employment in the global south, is unfortunately the second largest contributor to greenhouse gas emissions globally after energy. Farmers are vulnerable to climate-related problems such as droughts, floods, and crop failure due to extreme temperatures. To mitigate these problems, it is crucial to develop innovative technological solutions that cater to the specific climate and socio-economic needs of the agricultural sector, thereby enabling it to advance rapidly in developing countries. Identifying different land features, such as fields, trees, and dug wells, and analyzing them to optimize water consumption, crop yields, and soil carbon sequestration is essential. Therefore, we concentrate on three areas: Agricultural Landscape Understanding (ALU) for automatic identification of land features, Agriculture Monitoring and Event Detection (AMED) for automatic crop monitoring, and Soil Carbon Sequestration for a deeper understanding of the soil organic carbon change. DTSTART:20230304T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230303T120000Z UID:cb8c8fe56e3b4902db9d4ccba546d5d3-426 DTSTAMP:19700101T120011Z DESCRIPTION:Utilizing the CLT Structure in Stochastic Approximations of Sampling Algorithms URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/426/utilizing-the-clt-structure-in-stochastic-approximations-of-sampling-algorithms/ SUMMARY:We consider stochastic approximations of sampling algorithms like Langevin Monte Carlo and Interacting Particle Dynamics with random batches. These are heavily deployed in Bayesian inference, and the physical sciences. The noise induced by random batches is approximately Gaussian (due to the Central Limit Theorem) while the Brownian motion driving the algorithm is exactly Gaussian. We utilize this structure to provide improved guarantees for sampling algorithms under significantly weaker assumptions. We also propose covariance correction, which rescales the brownian motion to approximately remove the random batch error. We show that covariance corrected algorithms enjoy even better convergence. Joint work with: Aniket Das (Google) and Anant Raj (INRIA and UIUC) Speaker Website https://dheerajmn.mit.edu/ Microsoft Teams link https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rahul Madhavan, Rameesh Paul, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230303T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230303T120000Z UID:8133a997cb278e3a7a97df30ee6eb555-427 DTSTAMP:19700101T120011Z DESCRIPTION:Utilizing the CLT Structure in Stochastic Approximations of Sampling Algorithms URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/427/utilizing-the-clt-structure-in-stochastic-approximations-of-sampling-algorithms/ SUMMARY:We consider stochastic approximations of sampling algorithms like Langevin Monte Carlo and Interacting Particle Dynamics with random batches. These are heavily deployed in Bayesian inference, and the physical sciences. The noise induced by random batches is approximately Gaussian (due to the Central Limit Theorem) while the Brownian motion driving the algorithm is exactly Gaussian. We utilize this structure to provide improved guarantees for sampling algorithms under significantly weaker assumptions. We also propose covariance correction, which rescales the brownian motion to approximately remove the random batch error. We show that covariance corrected algorithms enjoy even better convergence. Joint work with: Aniket Das (Google) and Anant Raj (INRIA and UIUC) Speaker Website https://dheerajmn.mit.edu/ Microsoft Teams link https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rahul Madhavan, Rameesh Paul, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230303T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230309T120000Z UID:c46198dade0a93eb669a4fdf600081fe-428 DTSTAMP:19700101T120018Z DESCRIPTION:Causal Fairness Analysis URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/428/causal-fairness-analysis/ SUMMARY:Decision-making systems based on AI and machine learning have been used throughout a wide range of real-world scenarios, including healthcare, law enforcement, education, and finance. It is no longer far-fetched to envision a future where autonomous systems will drive entire business decisions and, more broadly, support large-scale decision-making infrastructure to solve societys most challenging problems. Issues of unfairness and discrimination are pervasive when decisions are being made by humans, and remain (or are potentially amplified) when decisions are made using machines with little transparency, accountability, and fairness. In this paper, we introduce a framework for *causal fairness analysis* with the intent of filling in this gap, i.e., understanding, modelling, and possibly solving issues of fairness in decision-making settings. The main insight of our approach will be to link the quantification of the disparities present in the observed data with the underlying, often unobserved, collection of causal mechanisms that generate the disparity in the first place, a challenge we call the Fundamental Problem of Causal Fairness Analysis (FPCFA). In order to solve the FPCFA, we study the problem of decomposing variations and empirical measures of fairness that attribute such variations to structural mechanisms and different units of the population. Our effort culminates in the Fairness Map, the first systematic attempt to organize and explain the relationship between various criteria found in the literature. Finally, we study which causal assumptions are minimally needed for performing causal fairness analysis and propose the Fairness Cookbook, which allows one to assess the existence of disparate impact and disparate treatment. <br> Joint work with Elias Bareinboim. The link to the paper is at https://causalai.net/r90.pdf. <br> Speaker Website: https://people.math.ethz.ch/~pleckod/ <br> Microsoft teams link: <br> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rahul Madhavan, Rameesh Paul, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230309T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230316T120000Z UID:961f14212b3047809a510016e440a92a-429 DTSTAMP:19700101T120011Z DESCRIPTION:Foundations of Lattice-based Cryptography URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/429/foundations-of-lattice-based-cryptography/ SUMMARY:Public key cryptography is essential for internet security, and RSA and Diffie-Hellman are the most widely used public-key cryptosystems for internet traffic. However, recent progress in building quantum computers threatens RSA and Diffie-Hellman's security, as they are vulnerable to quantum adversaries. To address this, organizations like the National Institute of Standards and Technology (NIST) and the European Telecommunications Standards Institute (ETSI) have started standardizing and deploying cryptosystems that are secure against quantum attacks. Recently, NIST has chosen Kyber and Dilithium, lattice-based candidates, as primary algorithms for security against quantum adversaries. The security of these cryptosystems crucially relies on the assumption that the best-known algorithms for the lattice problems cannot be significantly improved. <br> In this talk, I will discuss the connections between the security of lattice-based cryptosystems and the hardness of lattice problems. I will talk about classical and quantum algorithms for lattice problems. I will also discuss the works on the fine-grained security of lattice-based Crypto. DTSTART:20230316T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230328T120000Z UID:96286a5cc936c803abcf3abe2e851390-430 DTSTAMP:19700101T120016Z DESCRIPTION:The GRAMA project: Game theory, Random processes, Artificial intelligence, Machine learning for (Indian) Agriculture URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/430/the-grama-project-game-theory-random-processes-artificial-intelligence-machine-learning-for-indian-agriculture/ SUMMARY:We, at the Game Theory Lab in the Department of CSA, are currently engaged in a bouquet of projects exploring the use of game theory, optimization, and machine learning to address a few important problems in digital agriculture in the Indian context. These projects include: PREPARE, ACRE, CROPS, PROMISE, PROSPER, and AGRI-VAAHAN (acronyms will be expanded in the talk). PREPARE is concerned with crop price prediction; ACRE, with crop recommendation; CROPS, with crop planning; PROMISE, with procurement of agricultural inputs; and PROSPER, with markets for selling agricultural produce. AGRI-VAAHAN is an AIML pipeline for digital agriculture. In this talk, we provide an overview of these projects, some preliminary results, and work in progress. We present PROMISE in some detail highlighting the problems faced by the farmers in procuring quality inputs at affordable cost; we bring out how “farmer cooperatives” can creatively solve this problem with simple technology using standard tools from game theory and machine learning. These projects are supported by NABARD (National Bank for Agriculture and Rural Development) and BEL (Bharat Electronics Limited). DTSTART:20230328T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230316T120000Z UID:e5e9b0445e29a02820fcd0c6fd513412-432 DTSTAMP:19700101T120010Z DESCRIPTION:Private Convex Optimization via Exponential Mechanism URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/432/private-convex-optimization-via-exponential-mechanism/ SUMMARY:We study differentially private optimization of (non-smooth) convex functions F(x)=E_i[f_i(x)]. The classic exponential mechanism minimizes F(x) by sampling from pi(x) ~ exp(-kF(x)), but achieves a suboptimal privacy vs utility tradeoff. We show that modifying the exponential mechanism by adding an ell_2^2 regularizer to F(x) and sampling from pi(x) ~ exp(-k(F(x)+mu ||x||_2^2/2)) recovers both optimal empirical risk and population loss under (eps,delta)-DP. We also give an algorithm to efficiently sample from the exponential mechanism using optimal number of oracle queries to f_i(x). <br> We prove that the regularized exponential mechanism satisfies Gaussian Differential Privacy; our privacy bound is optimal (with tight constants), as it includes the analysis of Gaussian mechanism as a special case. The privacy proof uses isoperimetric inequality for strongly log-concave measures. <br> Joint work with Yin Tat Lee and Daogao Liu. The link to the paper is at https://arxiv.org/pdf/2203.00263.pdf. <br> <br> Speaker Website: https://www.microsoft.com/en-us/research/people/sigopi/ <br> <br> Microsoft teams link: <br> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d <br> <br> We are grateful to the Kirani family for generously supporting the theory seminar series <br> <br> Hosts: Rahul Madhavan, Rameesh Paul, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230316T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230320T120000Z UID:2cd5e752ab7c7c9989f8850afb981edd-433 DTSTAMP:19700101T120016Z DESCRIPTION:Criticality of AC0-formulae URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/433/criticality-of-ac0-formulae/ SUMMARY:Hastads celebrated switching lemma gives inverse-exponential bounds (In terms of t) on the probability that an k-DNF when hit by a p-restriction requires decision-trees of depth larger than t. The switching lemma has proved to be extremely powerful, since its discovery, leading to optimal size lower bounds for AC0-circuits [Hastad 1986] and AC0 formulae [Rossman 2015] against the parity function. <br> More recently, the search for optimal correlation bounds against parity led to the notion of criticality [Rossman 2019]. The criticality of a Boolean function f:{0,1}n → {0,1} is the minimum λ≥1 such that for all positive integers t, we have Prρ∼Rp[DTdepth(f|ρ)≥t]≤(pλ)t. <br> Hastad (2014) proved that size S and depth (d+1) AC0-circuits have criticiality at most O((logS)d) leading to optimal correlation bounds of AC0-circuits against parity. Rossman (2019) subsequently proved that size S and depth (d+1) AC0-formulae, which are regular (i.e., all gates of same depth have fan-in) having criticality at most O((logS/d)d). <br> In this work, we strengthen and unify all the above results by proving that any (not necessarily regular) AC0-formula of size S and depth (d+1) have criticality at most O((logS/d)d). The criticality bound implies tight correlation bounds against parity, tight Fourier concentration results and improved #SAT algorithm for AC0-formulae. <br> The talk is based on joint works with Tulasimohan Molli and Ashutosh Shankar. <br> Microsoft teams link: <br> Teams link <br> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d <br> <br> We are grateful to the Kirani family for generously supporting the theory seminar series <br> <br> Hosts: Rameesh Paul, Aditya Subramanian, Aditya Abhay Lonkar and Rahul Madhavan DTSTART:20230320T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230324T120000Z UID:463c5371e747a1eeafa6a566657290ec-434 DTSTAMP:19700101T120010Z DESCRIPTION:Astronomical Challenges in Atomic Scale Manufacturing URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/434/astronomical-challenges-in-atomic-scale-manufacturing/ SUMMARY:Semiconductor manufacturing is approaching the Angstrom-era with innovations such as gate all-around transistors, and chip-to-chip integration technologies enabling the continuation of Moores law. This talk will highlight some of the challenges that these advanced technologies pose to manufacturing semiconductors, and will cover how modern AI & HPC technologies are being leveraged to address these challenges to enable high-volume manufacturing. We will also give a peek into some of the solutions that KLA is pioneering in this space. DTSTART:20230324T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230324T120000Z UID:ac436da6a5946ea3d1474d8dcaf22545-435 DTSTAMP:19700101T120015Z DESCRIPTION:Cryptographic Primitives with Hinting Property URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/435/cryptographic-primitives-with-hinting-property/ SUMMARY:A hinting pseudorandom generator (PRG) is a potentially stronger variant of PRG with a ``deterministic`` form of circular security with respect to the seed of the PRG (introduced by Koppula and Waters in CRYPTO 2019). Hinting PRGs enable many cryptographic applications, most notably CCA-secure public-key encryption and trapdoor one-way functions. In this talk, I will cover a recent work where we study cryptographic primitives with the hinting property. Our work introduces a novel and conceptually simpler approach for designing hinting PRGs from certain decisional assumptions over cyclic groups or isogeny-based group actions, which enables simpler security proofs and new instantiations from concrete assumptions as compared to the existing approaches for designing such primitives. In this talk, I will present a detailed treatment of this simple approach for constructing hinting PRGs, including a concrete construction and proof from the DDH assumption over cyclic groups. Our work also introduces several extensions of hinting PRGs, such as: (i) a natural extension of the hinting property to weak pseudorandom functions (which we call hinting wPRFs),and (ii) a stronger version of the hinting property (which we call the functional hinting property) that guarantees security even in the presence of hints about functions of the secret seed/key. We show how to instantiate these extensions by building upon our simple approach to realize hinting PRGs, and also demonstrate that these extensions have stronger implications than plain hinting PRGs, particularly in realizing various notions of KDM-secure encryption. Additionally, we study the cryptographic complexity of hinting PRGs and show the first black-box separation between public-key encryption and hinting PRGs via a simple construction of hinting PRGs given only a random oracle (this black-box separation result also extends to hinting wPRFs). The talk will present a high-level overview of these results. Based on a joint work with Navid Alamati (https://eprint.iacr.org/2022/1770, appeared at ASIACRYPT 2022). Some prior background on cryptography will be useful, but not absolutely necessary. DTSTART:20230324T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230403T120000Z UID:f9065223526cf83ef4e1409c54a221f8-439 DTSTAMP:19700101T120010Z DESCRIPTION:Inducing Constraints in Paraphrase Generation and Consistency in Paraphrase Detection URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/439/inducing-constraints-in-paraphrase-generation-and-consistency-in-paraphrase-detection/ SUMMARY:Deep learning models typically require a large volume of data. Manual curation of datasets is time-consuming and limited by imagination. As a result, natural language generation (NLG) has been employed to automate the process. However, in their vanilla formulation, NLG model are prone to producing degenerate, uninteresting, and often hallucinated outputs [58]. Constrained generation aims to overcome these shortcomings by providing additional information to the generation process. Training data thus generated can help improve the robustness of deep learning models. Therefore, the central research question of the thesis is: <br> “How can we constrain generation models, especially in NLP, to produce meaningful outputs and utilize them for building better classification models?” <br> To demonstrate how generation models can be constrained, we present two approaches for paraphrase generation. Paraphrase generation involves the generation of text that conveys the same meaning as a reference text. We propose two strategies for paraphrase generation: <br> 1. DiPS (Diversity in Paraphrases using Submodularity): The first approach deals with constraining paraphrase generation to ensure d= iversity, i.e., ensuring that generated text(s) are sufficiently different from each other. We propose a decoding algorithm for obtaining diverse texts. We provide a novel formulation of the problem in terms of monotone submodular function maximization, specifically targeted toward the task of paraphrase generation. We demonstrate the effectiveness of our method for data augmentation on multiple tasks such as intent classification and paraphrase recognition. <br> 2. SGCP (Syntax Guided Controlled Paraphraser): The second approach deals with constraining paraphrase generation to ensure syntacticality, i.e., ensuring that the generated text is syntactically coherent with an exemplar sentence. We propose Syntax Guided Controlled Paraphraser (SGCP), an end-to-end framework for syntactic paraphrase generation without compromising relevance (fidelity). Through a battery of automated metrics and comprehensive human evaluation, we verify that this approach does better than prior works that utilize only limited syntactic information in the parse tree. <br> The second part (meaningful outputs) of the research question pertains to ensuring that the generated output is meaningful. Towards this, we present an approach for paraphrase detection to ascertain that the generated output is semantically coherent with the reference text. Paraphrase Detection is the task of detecting whether or not the two input natural language statements are paraphrases of each other. Fine-tuning pre-trained models such as BERT and RoBERTa on paraphrastic datasets have become the go-to approaches for such tasks. However, tasks like paraphrase detection are symmetric - they require the output to be invariant of the order of the inputs. In the traditional fine-tuned approach for paraphrase classification, inconsistency is often observed in the predicted labels or confidence scores based on the order of the inputs. We validate this shortcoming and apply a consistency loss function to alleviate inconsistency in symmetric classification. Our results show an improved consistency in predictions for three paraphrase detection datasets without a significant drop in the accuracy scores. <br> While these works address the research question via paraphrase generation and detection, the approaches presented here apply broadly to NLP-based deep learning models that require imposing constraints and ensuring consistency. The work on paraphrase generation can be extended to impose new kinds of constraints (for example, sentiment coherence) on generation, while paraphrase detection can be applied to ensure consistency in other symmetric classification tasks (for example, sarcasm interpretation) that use deep learning models. <br> <br> Microsoft teams link: <br> https://teams.microsoft.com/_#/l/meetup-join/19:meeting_MmIyZTdkYjYtOWFlMy00MTMwLWE4M2ItMDJhZjc1NThkYmQ5@thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2258c2a810-9762-463d-90d0-20832b3d1592%22%7d&anon=true&deeplinkId=c94373af-e3bc-4dd5-b297-35c5c191f0f5 DTSTART:20230403T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230411T120000Z UID:b2edc4436b22558a6e56cf3ed21ef32c-440 DTSTAMP:19700101T120014Z DESCRIPTION:Efficient Large Scale Model Training (part 2) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/440/efficient-large-scale-model-training-part-2/ SUMMARY:This will be a 2 part lecture series focusing on systems and techniques that enable efficient and reliable model training at scale. Training a deep neural network at scale requires a holistic use of all the datacenter resources including CPU, memory, storage, network, and the accelerators (GPUs). Large model training not only requires us to distribute the training process across several GPUs and possibly nodes, but also do so in the most efficient manner while avoiding data- and communication-stalls. The first part of the lecture will focus on techniques that enable multi-GPU and multi-node training of large models. These include, data, model, pipeline, and tensor parallelism and discuss scenarios where each of these techniques are useful. We will talk also touch upon the communication primitives involved in training, and how these distributed training techniques alleviate communication stalls. In the second part, we will discuss how to automatically parallelize DNN training by interleaving a subset of the techniques we have discussed based on the hardware and model characteristics. Finally, we will touch upon another important aspect of training efficiency, which is providing reliability via model checkpointing. Failures in hardware and software are inevitable during large model training, thereby necessitating low-cost checkpointing and training recovery. We will talk about why and how such reliable training can be achieved. Relevant reading : DistBelief [NeurIPS ‘12], Pipedream [SOSP ‘19], GPipe [NeurIPS ‘19], ALPA [OSDI’22], CheckFreq[FAST ‘21], CheckNRun [NSDI’22] DTSTART:20230411T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230406T120000Z UID:21c770c82b5df3d8d7785617de0f5eba-441 DTSTAMP:19700101T120014Z DESCRIPTION:Efficient Large Scale Model Training (part 1) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/441/efficient-large-scale-model-training-part-1/ SUMMARY:This will be a 2 part lecture series focusing on systems and techniques that enable efficient and reliable model training at scale. Training a deep neural network at scale requires a holistic use of all the datacenter resources including CPU, memory, storage, network, and the accelerators (GPUs). Large model training not only requires us to distribute the training process across several GPUs and possibly nodes, but also do so in the most efficient manner while avoiding data- and communication-stalls. The first part of the lecture will focus on techniques that enable multi-GPU and multi-node training of large models. These include, data, model, pipeline, and tensor parallelism and discuss scenarios where each of these techniques are useful. We will talk also touch upon the communication primitives involved in training, and how these distributed training techniques alleviate communication stalls. In the second part, we will discuss how to automatically parallelize DNN training by interleaving a subset of the techniques we have discussed based on the hardware and model characteristics. Finally, we will touch upon another important aspect of training efficiency, which is providing reliability via model checkpointing. Failures in hardware and software are inevitable during large model training, thereby necessitating low-cost checkpointing and training recovery. We will talk about why and how such reliable training can be achieved. Relevant reading : DistBelief [NeurIPS 12], Pipedream [SOSP 19], GPipe [NeurIPS 19], ALPA [OSDI 22], CheckFreq[FAST 21], CheckNRun [NSDI 22] DTSTART:20230406T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230405T120000Z UID:0342b8aff8a80efd8f8765539f44b00f-442 DTSTAMP:19700101T120016Z DESCRIPTION:Sequential learning in a stochastic multi armed bandit framework URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/442/sequential-learning-in-a-stochastic-multi-armed-bandit-framework/ SUMMARY:The classic stochastic multi armed bandit framework involves finitely many unknown probability distributions that can be sequentially sampled to generate independent rewards. In this talk we consider two foundational problems: First one corresponds to sampling to minimize the expected regret, or equivalently, to maximize the expected total reward. The second one corresponds to the best arm identification, i.e., identifying the arm with the largest mean, or any other performance measure, using as few samples as possible while providing explicit probabilistically correct selection guarantees. These problems form the bedrock of algorithms used in web design and advertising, recommendation systems, clinical trials and many other exciting applications. In this talk we review some of the popular algorithms used for these problems emphasizing the intuition underlying the elegant ideas. Technically speaking, these problems have been well studied under the restrictive assumption that arm distributions belong to a single parameter exponential family, that includes distributions such as Bernoulli and Gaussian with known variance. Under these settings, lower bounds on samples needed are developed using ideas from hypothesis testing, and algorithms are proposed that match the lower bound. We propose optimal algorithms that match the lower bounds even to a constant for general probability distributions under minimal restrictions. We further discuss how the proposed methodology leads to near optimal confidence intervals for distribution means. We discuss further enhancements in the presence of offline data that needs to be combined with online data. We further propose some new algorithms in the best arm identification setting that along with minimising sample complexity, are also computationally efficient. Speaker Website https://www.tcs.tifr.res.in/~sandeepj/ Organizers Note: The talk will be in two halves. We will have the first session from 16.00 to 17.00 followed by a short break for snacks, post which we have the second session. Microsoft Teams Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rahul Madhavan, Rameesh Paul, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230405T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230417T120000Z UID:2d0c9d2e213748a811c95a9e85966e86-443 DTSTAMP:19700101T120014Z DESCRIPTION:Stochastic Optimization And Its Application In Reinforcement Learning. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/443/stochastic-optimization-and-its-application-in-reinforcement-learning/ SUMMARY:Numerous engineering fields, including transportation systems, manufacturing, communication networks, healthcare, and finance, frequently encounter problems requiring optimization, including uncertainty. Simulation-based optimization is a workable substitute for accurate analytical solutions because of the numerous input variables and the need for a system model. Smoothed functional (SF) algorithms belong to the class of simultaneous perturbation methods that have been found useful for stochastic optimization problems particularly in high-dimensional parameter spaces. SF methods update the gradient of the objective using function measurements involving parameters that are perturbed simultaneously along all component directions. Katkovnik and Kulchitsky originally developed the SF gradient procedure. This results in the objective function getting smoothed because of the convolution. The objective function smoothing that results from the convolution with a smoothing density function can help the algorithm converge to a global minimum or to a point close to it. First we present a stochastic gradient algorithm for minimizing a smooth objective function that is an expectation over noisy cost samples, and only the latter are observed for any given parameter. Our algorithm employs a gradient estimation scheme with random perturbations, which are formed using the truncated Cauchy distribution from the $delta$ sphere. We analyze the bias and variance of the proposed gradient estimator. Our algorithm is found to be particularly useful in the case when the objective function is non-convex, and the parameter dimension is high. From an asymptotic convergence analysis, we establish that our algorithm converges almost surely to the set of stationary points of the objective function and obtain the asymptotic convergence rate. We also show that our algorithm avoids unstable equilibria, implying convergence to local minima. Further, we perform a non-asymptotic convergence analysis of our algorithm. In particular, we establish here a non-asymptotic bound for finding an $epsilon$-stationary point of the non-convex objective function. Finally, we demonstrate numerically through simulations that our algorithm outperforms GSF, SPSA and RDSA by a significant margin over a few non-convex settings and we further validate its performance over convex (noisy) objectives. Next we consider the problem of control in the setting of reinforcement learning (RL), where model information is not available. Policy gradient algorithms are a popular solution approach for this problem, and are usually shown to converge to a stationary point of the value function. We propose two policy Newton algorithms that incorporate cubic regularization. Both algorithms employ the likelihood ratio method to form estimates of the gradient and Hessian of the value function using sample trajectories. The first algorithm requires exact solution of the cubic regularized problem in each iteration, while the second algorithm employs an efficient gradient descent-based approximation to the cubic regularized problem. We establish convergence of our proposed algorithms to a second-order stationary point (SOSP) of the value function, which results in avoidance of traps in the form of saddle points. In particular, the sample complexity of our algorithms towards finding an $epsilon$-SOSP is $O(epsilon^{-3.5})$, and this is a significant improvement over the state-of-the-art sample complexity of $O(epsilon^{-4.5})$ DTSTART:20230417T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230411T120000Z UID:72af83eb83a2c50e121116ce19a49686-444 DTSTAMP:19700101T120016Z DESCRIPTION:Algorithms for packing and covering problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/444/algorithms-for-packing-and-covering-problems/ SUMMARY:We study fundamental geometric packing and covering problems and design algorithms with improved worst-case guarantees for them under various paradigms. In particular, we study the Strip Packing problem (SP), where we are given a vertical half-strip $[0,W]times[0,infty)$ and a set of $n$ axis-aligned rectangles of width at most $W$. The goal is to find a non-overlapping packing of all rectangles into the strip such that the height of the packing is minimized. A well-studied and frequently used practical constraint is to allow only those packings that are guillotine separable, i.e., every rectangle in the packing can be obtained by recursively applying a sequence of edge-to-edge axis-parallel cuts (guillotine cuts) that do not intersect any item of the solution. In this paper, we study approximation algorithms for the Guillotine Strip Packing problem (GSP), i.e., the Strip Packing problem where we require additionally that the packing needs to be guillotine separable. This problem generalizes the classical Bin Packing problem and also makespan minimization on identical machines, and thus it is already strongly $mathsf{NP}$-hard. Moreover, due to a reduction from the Partition problem, it is $mathsf{NP}$-hard to obtain a polynomial-time $(3/2-epsilon)$-approximation algorithm for GSP for any $epsilon>0$ (exactly as Strip Packing). We provide a matching polynomial time $(3/2+epsilon)$-approximation algorithm for GSP. Furthermore, we present a pseudo-polynomial time $(1+epsilon)$-approximation algorithm for GSP. This is surprising as it is $mathsf{NP}$-hard to obtain a $(5/4-epsilon)$-approximation algorithm for (general) Strip Packing in pseudo-polynomial time. Thus, our results essentially settle the approximability of GSP for both the polynomial and the pseudo-polynomial settings. In the context of covering, we study the Set Cover and the related dual Hitting Set problem, which are fundamental problems in combinatorial optimization. They are well-studied in the offline, online, and dynamic settings. In the offline version of set cover, $n$ elements from a universe $U$ and a set collection $mathcal{F}subseteq 2^U$ are given as input. The objective is to choose a minimum cardinality subcollection $mathcal{F}$ of $mathcal{F}$ such that all the elements are covered. This problem is known to be $mathsf{NP}$-hard to even approximate beyond $(1-epsilon)log n$ for a fixed $epsilon>0$. In the online version of set cover (resp. hitting set), $m$ sets (resp. $n$ points) are given and $n$ points (resp. $m$ sets) arrive online, one-by-one. In the dynamic versions, points (resp. sets) can arrive as well as depart. Our goal as before is to maintain a set cover (resp. hitting set), minimizing the size of the computed solution. We study the geometric versions of these problems, where the sets are geometric objects and the elements are points in a fixed euclidean space, and present new online and dynamic algorithms for them. For online set cover for axis-parallel squares of arbitrary sizes, we present a tight $O(log n)$-competitive algorithm, improving upon the $O(log nlog m)$ general case guarantee. In the same setting for hitting set, we provide a tight $O(log N)$-competitive algorithm, assuming that all points have integral coordinates in $[0,N)^{2}$. No online algorithm had been known for either of these settings, not even for unit squares (apart from the known online algorithms for arbitrary set systems). For both dynamic set cover and hitting set with $d$-dimensional hyperrectangles, we obtain $(log m)^{O(d)}$-approximation algorithms with $(log m)^{O(d)}$ worst-case update time. This partially answers an open question posed by Chan et al. [SODA22]. Previously, no dynamic algorithms with polylogarithmic update time were known even in the setting of squares (for either of these problems). Our main technical contributions are an extended quad-tree approach and a frequency reduction technique that reduces geometric set cover instances to instances of general set cover with bounded frequency. DTSTART:20230411T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230413T120000Z UID:3d385daa22c45dddf66f86d32f0783b3-445 DTSTAMP:19700101T120016Z DESCRIPTION:Data-free pruning of DNNs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/445/data-free-pruning-of-dnns/ SUMMARY:In this session there will be two talks on Data Free Pruning of DNNs, an emerging theme in Model Compression. Each talk is based on a paper to be presented at ICLR 2023. Talk 1: 4:00-4:30 Tanay Narshana Machine Learning Engineer Observe.AI Title: DFPC: Data flow driven pruning of coupled channels without data. Abstract: Deep Learning models have now become ubiquitous. However, the hardware requirements to deploy SoTA models are increasing at a faster rate than what Moores law can deliver. This makes such models challenging to deploy. Model compression, particularly pruning, is one way to alleviate this problem. Modern, multi-branched neural network architectures often possess complex interconnections like residual connections between layers, which we call coupled channels (CCs). Most existing works are typically designed for pruning single-branch models like VGG-nets. While these methods yield accurate subnetworks, the improvements in inference times when applied to multi-branch networks are comparatively modest. These methods do not prune CCs, which we observe contribute significantly to inference time. For instance, layers with CCs as input or output take more than 66% of the inference time in ResNet-50. Structured pruning of CCs in these multi-branch networks is an under-researched problem. Moreover, pruning in the data-free regime, where data is not used for pruning, is gaining traction owing to privacy concerns and computational costs associated with fine-tuning. In this talk, we present our recently accepted work at ICLR 2023 on the problem of pruning CCs in the data-free regime. The efficacy of our methodology is demonstrated via empirical results. We achieve up to 1.66x improvements in inference time for ResNet-101 trained on CIFAR-10 with a 5% accuracy drop without fine-tuning. With access to the ImageNet training set, we achieve significant improvements over the data-free method and see an improvement of at least 47.1% in speedup for a 2.3% accuracy drop for ResNet-50 against our baselines. Talk2: 4:30-5:00 Chaitanya Murti PhD Student Robert Bosch Centre for Cyberphysical Systems IISc Title: TVSPrune - Purning Nondescriminative filters Total Variation Separability of Intermediate Features Abstract: Achieving structured, data-free sparsity of deep neural networks (DNNs) remains an open area of research. In this work, a solution to the problem of pruning filters with only access to the original data distribution, and without access to the original training set or loss function is proposed. The solution is based on the following hypothesis:well-trained models possess discriminative filters, and any non-discriminative filters can be pruned without impacting the predictive performance of the classifier. A new paradigm for pruning neural networks is proposed based on this hypothesis: distributional pruning, wherein access to the distributions that generated the original datasets is required. The discriminative ability of filters is formalised and quantified using the total variation (TV) distance between the class-conditional distributions of the filter outputs. Next, the LDIFF score is proposed. The LDIFF score is a heuristic to quantify the extent to which a layer possesses a mixture of discriminative and non-discriminative filters. The main contribution is a novel one-shot pruning algorithm, called TVSPrune, that identifies non-discriminative filters for pruning. This algorithm is extended to IterTVSPrune, wherein TVSPrune is applied iteratively, thereby enabling greater sparsification of a given model. DTSTART:20230413T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230414T120000Z UID:57b1e74c65404b70a2d8fd91148cc1c9-446 DTSTAMP:19700101T120011Z DESCRIPTION:Hop-Constrained Expander Decompositions, Oblivious Routing, and Distributed Universal Optimality URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/446/hop-constrained-expander-decompositions-oblivious-routing-and-distributed-universal-optimality/ SUMMARY:In a recent result Häupler, Ghaffari, and Zuzic [STOC 2021] showed that so-called h -hop oblivious routing schemes with a polylogarithmic competitive ratio exist. These can be used to solve packet routing problems almost optimally, i.e., with only a polylogarithmic loss in routing time compared to a globally optimal solution. In this talk we present a different way of constructing h -hop oblivious routing schemes that have a weaker (sub-polynomial) competitive ratio but which can be constructed very efficiently by a distributed algorithm. This result has important consequences in the area of distributed computing: it gives novel CONGEST algorithms for a large class of important optimization problems, including minimum-spanning tree, ( 1 + ϵ )-min-cut, ( 1 + ϵ )-shortest paths. Our algorithms solve these problems in sub-polynomial rounds on any network, as long as a sub-polynomial-round distributed algorithm exists for this network. This is joint work with Bernhard Häupler and Mohsen Ghaffari. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rameesh Paul, Rahul Madhavan, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230414T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230414T120000Z UID:6d37a83dc2fbd4aa097bf7b24a072903-447 DTSTAMP:19700101T120014Z DESCRIPTION:How to Play with Witness Encryption without the Theoretical Hassle (Or: weak forms of WE and good stuff we can get from them) URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/447/how-to-play-with-witness-encryption-without-the-theoretical-hassle-or-weak-forms-of-we-and-good-stuff-we-can-get-from-them/ SUMMARY:Witness encryption (WE), allows one to encrypt a message to a statement for some NP language, such that any user holding a witness for it can decrypt the ciphertext. If we could construct this primitive, we would be able to do without certificate authorities, but also to use it as an extremely versatile building block in other cryptographic applications. Unfortunately, from a theoretical standpoint, it is still unclear whether we will be able to instantiate a general-purpose witness encryption scheme from reliable assumptions soon (unless we go through iO). In this talk we tackle the questions: What are other weak-but-useful variants of WE that we can actually construct? And what efficiency properties would we require from them? We discuss some recent works in this direction and their applications, in particular on forms of non-interactive (and reusable) MPC (Ben and Lin, TCC20), where parties can securely compute a function by broadcasting a single message, assuming only an encoding of their input exists on a bulletin board. This talk is partly a presentation of the work in: https://eprint.iacr.org/2022/1510.pdf DTSTART:20230414T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230428T120000Z UID:34528e7a42651bd37d05ec61424c5610-448 DTSTAMP:19700101T120016Z DESCRIPTION:On symmetries of and equivalence tests for two polynomial families and a circuit class URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/448/on-symmetries-of-and-equivalence-tests-for-two-polynomial-families-and-a-circuit-class/ SUMMARY:Two n-variate polynomials f and g in F[x1,...,xn] are said to be equivalent over the field F if there exists an invertible matrix A over F such that f = g(Ax), where x = (x1 … xn). The problem of testing whether f is equivalent to a polynomial g coming from a polynomial family G (or computed by a circuit C in a circuit class D) is called the equivalence test (in short, ET) for G (respectively, D). In this thesis, we study equivalence tests for the determinant polynomial family and the class of regular read-once arithmetic formulas (ROFs). We also study some structural and algorithmic properties related to the symmetries of the Nisan-Wigderson design polynomial (in short, NW) and solve an interesting special case of ET for NW. In the first work, we study ET for the determinant (in short, DET) over finite fields and the field of rational numbers denoted Q. A randomized polynomial time DET over the field of complex numbers was given by Kayal in [Kay12]. But DET over finite fields and over Q were not known. We give the first randomized polynomial-time DET over finite fields and also give the first DET over Q. The DET over Q takes oracle access to an integer factoring algorithm (IntFact) and if the input polynomial f is equivalent to the n x n determinant over Q, then it outputs a certificate matrix A over Q. This algorithm is randomized and is efficient for bounded values of n. Assuming the generalized Riemann hypothesis, we also show that the problem of integer factoring reduces to DET for quadratic forms (i.e., n = 2 case). We also give another DET over Q, which does not require oracle access to IntFact, but it outputs a certificate matrix over an extension field L of Q, where the degree of this extension is at most n. This DET algorithm is also randomized and is efficient for any value of n. The DET algorithms over finite fields and Q are obtained by decomposing the Lie algebra of f and then invoking known algorithms for the full matrix algebra isomorphism (FMAI) problem over finite fields and Q. FMAI is a well-studied problem in computer algebra. We also give a reduction from FMAI to DET, which is efficient when n is bounded. This is joint work with Ankit Garg, Neeraj Kayal, and Chandan Saha. In the second work, we study ET for read-once arithmetic formulas (ROFs). An ROF is an arithmetic formula where every leaf node is labeled by either a distinct variable or a constant from the underlying field. ROFs are well-studied in the literature. In this work, we give the first randomized polynomial-time ET with oracle access to quadratic form equivalence for certain restricted ROFs, which we call regular ROFs. ET for regular ROFs generalizes the well-known quadratic form equivalence problem over the field of complex numbers and ETs for the classes of sum-product polynomials and ROANFs. ETs for these two classes have been recently studied by Medini & Shpilka (2021). Our ET algorithm uses some crucial properties related to the non-zeroness, the factors, and essential variables of the Hessian determinant of a regular ROF. We study these properties for the Hessian determinant of an arbitrary ROF C by analyzing the structures and coefficients of some nice monomials in the Hessian determinant of C. This is joint work with Chandan Saha and Bhargav Thankey. In the last work, we study some structural and algorithmic properties related to the symmetries of NW and give a special case of ET for NW. In NW, each pair of monomials has very few variables in common. This property of NW has been exploited to give strong lower bounds for different classes of arithmetic circuits. Like NW, other polynomials like the permanent, the determinant, etc., have also been extensively used in lower bound results. But unlike these polynomials, not much is known about NW. In this work, we study some important properties of NW related to its symmetries. A matrix A is said to be a symmetry of NW if NW(Ax) = NW. We show that NW is characterized by its symmetries over the field of complex numbers but not over the fields of real numbers and rational numbers. Using the symmetries of NW, we show that NW is characterized by circuit identities over any field. This result implies a randomized polynomial time circuit testing algorithm for NW - which tests whether some circuit C computes NW- and the flip theorem for NW. We also solve an interesting special case of ET for NW, which we call the block-diagonal permutation scaling ET for NW. This ET uses the symmetries of NW crucially. This is joint work with Chandan Saha. Microsoft Teams Link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjM0NjkzY2QtYTNmNS00M2ZlLTk0YjYtNjU1ZmE4NGEwMThi%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%222adc8ab7-d96f-4eb9-bd5e-46f77c3201c6%22%7d DTSTART:20230428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230424T120000Z UID:b419ea45c2e1e951c27e7584999cbae0-449 DTSTAMP:19700101T120010Z DESCRIPTION:Privadome: A System for Citizen Privacy in the Delivery Drone Era URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/449/privadome-a-system-for-citizen-privacy-in-the-delivery-drone-era/ SUMMARY:E-commerce companies are actively considering the use of delivery drones for customer fulfillment, leading to growing concerns around citizen privacy. Drones are equipped with cameras, and the video feed from these cameras is often required as part of routine navigation, be it for semi-autonomous or fully-autonomous drones. Footage of ground based citizens captured in these videos may lead to privacy concerns. This M.Tech. (Research) thesis presents the design, implementation and evaluation of Privadome, a system that implements the vision of a virtual privacy dome centered around the citizen. Privadome is designed to be integrated with city-scale regulatory authorities that oversee delivery drone operations and realizes this vision through two components, Pd-Mpc and Pd-Ros. Pd-Mpc allows citizens equipped with a mobile device to identify drones that have captured their footage. It uses secure two-party computation to achieve this goal without compromising the privacy of the citizens location. Pd-Ros allows the citizen to communicate with such drones and obtain an audit trail showing how the drone uses their footage and determine if privacy-preserving steps are taken to sanitize the footage. An experimental evaluation of Privadome shows that the system scales to near-term city-scale delivery drone deployments (hundreds of drones). We show that with Pd-Mpc the mobile data usage on the citizens mobile device is comparable to that of routine activities on the device, such as streaming videos. We also show that the workflow of Pd-Ros consumes a modest amount of additional CPU resources and power on our experimental platform. DTSTART:20230424T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230502T120000Z UID:203d81b672c3a8c3e874014d25930afa-450 DTSTAMP:19700101T120011Z DESCRIPTION:Abstractions for Network Control Plane Verification URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/450/abstractions-for-network-control-plane-verification/ SUMMARY:The network control plane is a complex distributed system that runs various protocols for exchanging messages between routers and selecting paths for routing traffic. Errors in control plane configurations can lead to expensive outages or critical security breaches, leading to great interest in applying formal methods to ensure correctness. Although verification approaches based on use of Satisfiability Modulo Theory (SMT) solvers are general and powerful, they face scalability challenges. I will describe our recent work on key abstractions and modular assume-guarantee reasoning that have enabled our SMT-based approach to successfully handle large-sized networks (with several thousands of routers), similar to those in operation in modern data centers. This talk describes joint work with Ryan Beckett, Ratul Mahajan, Divya Raghunathan, Timothy Alberdingk Thijm, and David Walker. DTSTART:20230502T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230508T120000Z UID:1543b9e2002a0c0df404da763011cfab-452 DTSTAMP:19700101T120011Z DESCRIPTION:Secure Computation Protocol Suite for Privacy-Conscious Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/452/secure-computation-protocol-suite-for-privacy-conscious-applications/ SUMMARY:As an alternative to performing analytics in the clear, there is an increasing demand for developing privacy-preserving solutions that aim to protect sensitive data while still allowing for its efficient analysis. Among the various privacy-enhancing technologies, secure multiparty computation (MPC) is a promising approach that enables multiple parties (n) to jointly process their private inputs while ensuring that no coalition of at most t < n parties, under the control of an adversary, learns any information other than the intended output. In this thesis, we identify various such real-world applications that demand privacy-preserving solutions and address these via MPC. We consider a broad range of applications that span across healthcare, finance, and even social sectors. For each application under consideration, we identify the desirable MPC setting (e.g., number of computing parties, n) and security notion to be achieved when designing the solution. Based on this, we either design new MPC frameworks that provide improved security guarantees and efficiency or enhance the existing frameworks. <br> Although we make application-specific design choices, the common theme while designing secure protocols for all is to design as efficient a solution as possible. In this regard, we make the following common design choices across all applications. First, we consider an honest majority among the computing parties, which is known to render efficient protocols in comparison to the dishonest majority. Second, we focus on designing secure protocols in the preprocessing paradigm, where expensive input-independent computations are pushed onto a preprocessing phase, thereby making way for a fast and efficient input-dependent online phase. Finally, our protocols are designed to operate on the ring algebraic structure to capitalize on the efficiency gains obtained from utilizing the CPU architecture. We next elaborate on the specific applications considered in the thesis and the contributions therein. <br> <br> Secure computation over graphs via traditional security notion. <br> Operating on graph-structured data is ubiquitous due to the modelling capabilities of graphs, and this finds use in analyzing various systems like social networks, biological networks, transportation networks, etc. However, privacy concerns arise when analyzing graphs that model sensitive data. To address this, we design privacy-preserving solutions for two popular graph algorithms---local clustering and graph convolutional networks. <br> 1. Secure local clustering: Identifying a cluster around a target node in a graph, termed local clustering, finds use in several applications, including fraud detection, targeted advertising, community detection, etc. We design solutions for privacy-preserving local clustering, which is done for the first time in the literature. Keeping efficiency in mind for large graphs, we build over the best-known honest-majority 3-party framework of SWIFT (USENIX 21) and enhance it with some of the necessary yet missing primitives. To further enhance efficiency, we design the protocols using the GraphSC paradigm, which provides a generic secure framework for efficiently evaluating graph algorithms. Since this paradigm relies on a secure shuffle primitive, we also design an efficient secure 3-party shuffle protocol. We note that secure shuffle is a versatile primitive that finds widespread use in various other applications as well (which may not involve computations over a graph), such as electronic voting, oblivious RAM, and anonymous broadcast, to name a few. Hence, as a by-product of our shuffle protocol, we are also able to securely realize an anonymous broadcast system. As the name suggests, anonymous broadcast enables a set of N clients to anonymously broadcast their messages while guaranteeing that none learns about the association between a message and the identity of its sender. Hence, while anonymous broadcast may not be inherently associated with graph computations, we diverge slightly to demonstrate how our shuffle protocol can be employed to realize anonymous broadcast in the 3-party setting, as considered in prior works. In the process, not only do we design a more efficient anonymous broadcast system compared to the state-of-the-art, but our system also provides improved security guarantees and properties such as censorship resistance that were missing in the prior solution. <br> 2. Secure graph convolutional networks: Graph convolutional networks (GCNs) are gaining popularity due to their ability to effectively model and learn from complex graph-structured data. We put forth Entrada, a framework for securely evaluating GCNs. For efficiency and accuracy reasons, Entrada builds over the 4-party framework of Tetrad (NDSS 22) and enhances the same by providing the necessary primitives. Moreover, Entrada leverages the GraphSC paradigm to further enhance efficiency and entails designing a secure and efficient shuffle protocol specifically in the 4-party setting. This, to the best of our knowledge, is done for the first time and may be of independent interest. <br> Stepping beyond traditional security for financially and socially relevant problems. Most protocols in the small-party setting that are designed to attain the strongest security notion of guaranteed output delivery (GOD), rely on entrusting an honest party, identified as the trusted third party (TTP), with inputs in the clear to carry out the computation. However, this may not be desirable for certain applications that deal with highly sensitive data. Another drawback of traditional MPC protocols is the view leakage attack, where a malicious adversary may send its view to an honest party, thereby enabling the latter to obtain the underlying secret information. To address these drawbacks in the traditional MPC definition, Alon et al.(CRYPTO 20) propose the notion of MPC with Friends and Foes (FaF). Thus, departing from the traditional MPC model, we identify the need to design FaF-secure MPC protocols for applications that deal with highly sensitive information, where information leakage must be prevented even against quorums of honest parties. Specifically, we consider the applications of secure dark pools and secure allegation escrow systems. Keeping efficiency at the centre stage, we design FaF-secure 5-party computation protocols (5PC) that consider one malicious and one semi-honest corruption and constitute the optimal setting for attaining an honest majority. <br> 1. Secure dark pools: Dark pools are private security exchanges that allow investors to trade financial instruments outside of the prying eyes of the public and ensure the trade remains unexposed until it is completed. Dark pools are traditionally operated by centralized trusted brokers, who, in the past, have been known to misuse insider information. This necessitates designing solutions that guarantee privacy even against the dark pool operator. Hence, given the sensitive nature of financial data that is involved in the computation and the drawbacks present in the traditional MPC solutions, we design FaF secure solutions for the same in the 5PC setting. We design improved solutions for the continuous double auction (CDA) and volume-based matching algorithms that are used in dark pools. We benchmark the performance of these secure matching algorithms and observe improvements in comparison to the prior works. <br> <br> 2. Secure allegation escrow system: The rising issues of malpractices have led victims to seek comfort by acting in unison against common perpetrators (e.g., #MeToo movement). To increase trust in the system, cryptographic solutions are being designed to realize secure allegation escrow (SAE) systems. In this regard, we identify privacy issues present in prior works and put forth an SAE system to arrest all these breaches. Given the highly sensitive nature of allegation data, we choose to realize the system under FaF security as opposed to traditional security notions. We also provide additional features which were absent in the state of the art. We benchmark the proposed system with the FaF secure 5PC protocols to showcase the practicality of our solution. <br> Secure computation with a constant number of parties. <br> Unlike the applications considered above that demanded operating with a specific number of parties, the latter may vary depending on the application. Hence, we provide a generalization which allows instantiating the MPC protocol with an arbitrary (constant) number of parties (n). Our generalized protocols continue to operate in the honest majority setting to capitalize on the efficiency benefits that this setting provides over the dishonest majority, which thereby facilitates attaining an efficient solution for the end application. We design two different protocols, that are secure against a semi-honest and a malicious adversary, respectively. We also design a wide range of building blocks that facilitate the secure realization of various applications, including but not limited to genome sequence matching, biometric matching, and even deep neural networks, and showcase the practicality of the designed protocols by benchmarking these applications. <br> In this way, we design a range of building blocks in various MPC settings that can facilitate secure realizations for the above-mentioned privacy-conscious applications. <br> References: <br> [1] Nishat Koti, Varsha Bhat Kukkala, Arpita Patra, Bhavish Raj Gopal. “Shield: Secure Allegation Escrow System with Stronger Guarantees.” TheWebConf, 2023.<br> [2] Pranav Shriram A, Nishat Koti, Varsha Bhat Kukkala, Arpita Patra, Bhavish Raj Gopal, Somya Sangal. “Ruffle: Rapid 3-Party Shuffle Protocols.” PoPETS, 2023.<br> [3] Pranav Shriram A, Nishat Koti, Varsha Bhat Kukkala, Arpita Patra, Bhavish Raj Gopal. “Find Thy Neighbourhood: Privacy-Preserving Local Clustering.” PoPETS, 2023.<br> [4] Nishat Koti, Varsha Bhat Kukkala, Arpita Patra, Bhavish Raj Gopal. “PentaGOD: Stepping beyond Traditional GOD with Five Parties.” CCS, 2022.<br> [5] Nishat Koti, Shravani Patil, Arpita Patra, Ajith Suresh. “MPClan: Protocol Suite for Privacy-Conscious Computations.” Under submission.<br> [6] Nishat Koti, Varsha Bhat Kukkala, Arpita Patra, Bhavish Raj Gopal. “Entrada to Secure Graph Convolutional Networks for Defying Fraud.” Under submission. DTSTART:20230508T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230428T120000Z UID:46bed363dc37a286ca80decced8cde61-453 DTSTAMP:19700101T120016Z DESCRIPTION:Online and Bandit Algorithms Beyond â„“_p Norms URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/453/online-and-bandit-algorithms-beyond-a_p-norms/ SUMMARY:Vector norms play a fundamental role in computer science and optimization, so there is an ongoing effort to generalize existing algorithms to settings beyond â„“_∞ and â„“_p norms. We show that many online and bandit applications for general norms admit good algorithms as long as the norm can be approximated by a function that is “gradient-stable”, a notion that we introduce. Roughly it says that the gradient of the function should not drastically decrease (multiplicatively) in any component as we increase the input vector. We prove that several families of norms, including all monotone symmetric norms, admit a gradient-stable approximation, giving us the first online and bandit algorithms for these norm families. In particular, our notion of gradient-stability gives O(log^2(dimension))-competitive algorithms for the symmetric norm generalizations of Online Generalized Load Balancing and Bandits with Knapsacks. Our techniques extend to applications beyond symmetric norms as well, e.g., to Online Vector Scheduling and to Online Generalized Assignment with Convex Costs. Some key properties underlying our applications that are implied by gradient-stable approximations are a “smooth game inequality” and an approximate converse to Jensens inequality. Joint work with Marco Molinaro and Sahil Singla Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rameesh Paul, Rahul Madhavan, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230428T120000Z UID:3d89f571d920d4ceb29fce7b8fa442a8-455 DTSTAMP:19700101T120011Z DESCRIPTION:Improved Approximation Bounds On Maximum Edge q-coloring Of Dense Graphs URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/455/improved-approximation-bounds-on-maximum-edge-q-coloring-of-dense-graphs/ SUMMARY:The anti-Ramsey number ar(G,H) with input graph G and pattern graph H, is the maximum positive integer k such that there exists an edge coloring of G using k colors, in which there are no rainbow subgraphs isomorphic to H in G. (H is rainbow if all its edges get distinct colors). The concept of anti-Ramsey number was introduced by Erdos, Simanovitz, and Sos in 1973. Thereafter several researchers investigated this concept in the combinatorial setting. The cases where pattern graph H is a complete graph K_r, a path P_r or a star K_{1,r} for a fixed positive integer r, are well studied. <br> Recently, Feng et al. revisited the anti-Ramsey problem for the pattern graph K_{1,t} (for t geq 3) purely from an algorithmic point of view, due to its applications in interference modeling of wireless networks. They posed it as an optimization problem, the maximum edge q-coloring problem. For a graph G and an integer q geq 2, an edge q-coloring of G is an assignment of colors to edges of G, such that edges incident on a vertex span at most q distinct colors. The maximum edge q-coloring problem seeks to maximize the number of colors in an edge q-coloring of the graph G. Note that the optimum value of the edge q-coloring problem of G equals ar(G,K_{1,q+1}). <br> We study ar(G,K_{1,t}), the anti-Ramsey number of stars, for each fixed integer t geq 3, both from combinatorial and algorithmic point of view. The first of our main results, presents an upper bound for ar(G,K_{1,q+1}), in terms of number of vertices and the minimum degree of G. The second one improves this result for the case of triangle free input graphs. <br> For a positive integer t, let H_t denote a subgraph of G with maximum number of possible edges and maximum degree t. From an observation of Erdos, Simanovitz, and Sos, we get: |E(H_{q-1})| + 1 leq ar(G,K_{1,q+1}) leq |E(H_{q})|. For instance, when q=2, the subgraph E(H_{q-1}) refers to a maximum matching. It looks like |E(H_{q-1})| is the most natural parameter associated with the anti-ramsey number ar(G,K_{1,q+1}) and the approximation algorithms for the maximum edge coloring problem proceed usually by first computing the H_{q-1}, then coloring all its edges with different colors and by giving one (sometimes more than one) extra colors to the remaining edges. The approximation guarantees of these algorithms usually depend on upper bounds for ar(G,K_{1,q+1}) in terms of |E(H_{q-1})|. <br> Our third main result presents an upper bound for ar(G,K_{1,q+1}) in terms of |E(H_{q-1})|. <br> All our results have algorithmic consequences. For some large special classes of graphs, such as d-regular graphs, where d geq 4, our results can be used to prove a better approximation guarantee for the sub-factor based algorithm. We also show that all our bounds are almost tight. <br> Results for the case q=2 were done earlier by Chandran et al. In this thesis, we extend it further for each fixed integer q greater then 2. <br> Microsoft teams link: <br> https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZWIyOWQyNTgtN2Y3Mi00NzYyLWI1YWYtZTQ0ZDhiNThiMTI4%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22d59009b4-da58-4d38-a9a5-6ed4fc892dd0%22%7d DTSTART:20230428T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230504T120000Z UID:bf24072bcde2c3b9fdec8295457abcf8-456 DTSTAMP:19700101T120015Z DESCRIPTION:Realizing quantum advantage across multiple applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/456/realizing-quantum-advantage-across-multiple-applications/ SUMMARY:Quantum computing offers a new path towards solving problems well beyond the scope of the most powerful classical computers. This is popularly referred to as DTSTART:20230504T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230504T120000Z UID:a8690babe52e42f4ff7d437b60b14916-457 DTSTAMP:19700101T120016Z DESCRIPTION:A PTAS for Unsplittable Flow on a Path URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/457/a-ptas-for-unsplittable-flow-on-a-path/ SUMMARY:In the Unsplittable Flow on a Path problem (UFP) we are given a path with edge capacities, and a set of tasks where each task is characterized by a subpath, a demand, and a weight. The goal is to select a subset of tasks of maximum total weight such that the total demand of the selected tasks using each edge e is at most the capacity of e. The problem admits a QPTAS [Bansal, Chakrabarti, Epstein, Schieber, STOC06; Batra, Garg, Kumar, Mömke, Wiese, SODA15]. After a long sequence of improvements [Bansal, Friggstad, Khandekar, Salavatipour, SODA09; Bonsma, Schulz, Wiese, FOCS11; Anagnostopoulos, Grandoni, Leonardi, Wiese, SODA14; Grandoni, Mömke, Wiese, Zhou, STOC18], the best known polynomial time approximation algorithm for UFP has an approximation ratio of 1+1/(e+1) + ε < 1.269 [Grandoni, Mömke, Wiese, SODA22]. It has been an open question whether this problem admits a PTAS. We solve this open question and present a polynomial time (1 + ε)-approximation algorithm for UFP. Microsoft Teams links: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rameesh Paul, Rahul Madhavan, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230504T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230508T120000Z UID:f3434ddc6fb098e442cceb1c404faec6-458 DTSTAMP:19700101T120016Z DESCRIPTION:Handling competitive aspects of Synchronization in Shared Environments using lock usage fairness URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/458/handling-competitive-aspects-of-synchronization-in-shared-environments-using-lock-usage-fairness/ SUMMARY:In shared environments such as operating systems, multiple tenants with varied requirements compete to access the shared resources, making strong performance isolation necessary. Locks are one of the widely used synchronization primitives that provide mutual exclusion in such environments. In this talk, I will emphasize the competitive aspects of synchronization in such shared environments. I will start by introducing the notion of lock usage -- a new lock property that deals with the time spent in the critical section. Then, I will show how unfair lock usage in shared environments leads to two new problems -- scheduler subversion and adversarial synchronization. I will then introduce two solutions -- Scheduler Cooperative Locks and Tratr that view synchronization as a resource to mitigate these two problems. Lastly, I will talk about our current work on ensuring hierarchical fairness using lock usage and the design of work-conservative Scheduler Cooperative Locks. DTSTART:20230508T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230512T120000Z UID:eec275e1ac61151006df68791f52180f-459 DTSTAMP:19700101T120014Z DESCRIPTION:CodeQueries: Benchmarking Query Answering over Source Code URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/459/codequeries-benchmarking-query-answering-over-source-code/ SUMMARY:Software developers often make queries about the security, performance effectiveness, and maintainability of their code. Through an iterative debugging process, developers analyze the code to find answers to these queries. The process can be seen as a question-answering task that requires developers to identify code spans satisfying certain properties. Many of these queries can be answered by existing code analysis tools such as CodeQL. However, using such tools requires design, implementation, and verification efforts. In this work, we propose an alternative to the code analysis tools by formulating the task of query answering over source code as a span prediction problem. In the proposed approach, a neural model is designed to predict appropriate answer spans in a code in response to a query. The required supporting-facts to justify the predicted answers are also identified by the model. Pre-trained language models for code are fine-tuned on a newly prepared challenging dataset, CodeQueries, for query answering over source code. We demonstrate that the proposed approach performs well on the query answering over source code task when only relevant code blocks are provided as input to the model. Experiments conducted on the dataset demonstrate that the proposed neural approach is robust to noisy span labeling and can even handle code with minor syntax errors. Although large-sized code and limited training examples adversely affect the model performance, we suggest methods to address these issues. Based on our study, we believe that the proposed neural approach will be an additional tool in a developers toolbox for query answering over source code. DTSTART:20230512T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230512T120000Z UID:d40a8407258538e250e9fdf7a1ff1ab8-460 DTSTAMP:19700101T120011Z DESCRIPTION:Instance-dependent Sample Complexity Bounds for Zero-sum Matrix Games URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/460/instance-dependent-sample-complexity-bounds-for-zero-sum-matrix-games/ SUMMARY:We study the sample complexity of identifying an approximate equilibrium for two-player zero-sum n×2 matrix games. That is, in a sequence of repeated game plays, how many rounds must the two players play before reaching an approximate equilibrium (e.g., Nash)? We derive instance-dependent bounds that define an ordering over game matrices that captures the intuition that the dynamics of some games converge faster than others. Specifically, we consider a stochastic observation model such that when the two players choose actions i and j, respectively, they both observe each others played actions and a stochastic observation Xij such that EXij = Aij. To our knowledge, our work is the first case of instance-dependent lower bounds on the number of rounds the players must play before reaching an approximate equilibrium in the sense that the number of rounds depends on the specific properties of the game matrix A as well as the desired accuracy. We also prove a converse statement: there exist player strategies that achieve this lower bound <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d </a> <br> We are grateful to the Kirani family for generously supporting the theory seminar series <br> <br> Hosts: Rameesh Paul, Rahul Madhavan, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230512T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230515T120000Z UID:a9e660a9afbdd1a57fc192d55155ea87-461 DTSTAMP:19700101T120015Z DESCRIPTION:Towards Self-Sustainable Wearable Devices for Reliable Mobile Health Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/461/towards-self-sustainable-wearable-devices-for-reliable-mobile-health-applications/ SUMMARY:Wearable sensors, along with smart home technologies, have the potential to transform healthcare by enabling cost-effective, reliable, continuous, and data-driven monitoring of users in a free-living environment. Continuous monitoring of user activities and vital signs is exciting, as it can expose health issues otherwise difficult to perceive in an office visit. It can also identify early warnings of adverse health events requiring immediate intervention. Despite the impressive potential of wearable technology, widespread adoption of wearable devices has been limited due to several technology and adaptation challenges. This has led to multiple health societies, including the Movement Disorders Society Task Force on Technology, stating that solving these challenges is crucial to improving wearable devices adoption. First, existing wearables devices are rigid, which makes them uncomfortable to wear for long periods. Second, existing implementations of wearable devices typically consider a single device or sensor at a time. As a result, the development of wearable devices happens in isolated silos with limited cross-compatibility. Third, wearable devices have small batteries that necessitate frequent recharging to prolong their operation. In this talk, I will present solutions towards these challenges. First, I will present our open-source hardware/software platform for wearable health monitoring. The platform uses flexible hybrid electronics to enable devices that conform to the shape of the users body. Then, I will describe an algorithm to enable recharge-free operation of wearable devices that harvest energy from the environment. After that, I will present a hardware accelerator and reliable algorithms for the human activity recognition application. DTSTART:20230515T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230519T120000Z UID:e7c3da5e8e16412eda7b16e320c27da1-462 DTSTAMP:19700101T120016Z DESCRIPTION:Algorithms for Individual and Collective Fairness Measures URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/462/algorithms-for-individual-and-collective-fairness-measures/ SUMMARY:The problem of fair allocation has been a central topic in economic theory, and the literature on fair division has provided fundamental insights on how to allocate resources among agents in a fair manner. By drawing upon existing literature, this thesis focuses on computational challenges that arise in different settings of fair-division problems. The thesis presents efficient algorithms, including approximation algorithms where applicable, for fair resource allocation by optimizing for different types of fairness measures. We also complement these algorithms by providing matching hardness results demonstrating the tightness of the obtained approximation guarantees. The algorithms presented in this thesis provide new tools to address fair allocation problems in practice and offer insights into the design of efficient procedures for resource allocation. This thesis is structured into two parts, each focusing on a distinct type of fairness measure: collective criteria and individual criteria. Part-I: Collective Fairness Algorithms for maximizing p-mean welfare We propose a polynomial-time algorithm for allocating indivisible goods among agents with subadditive valuations. We consider p-mean welfare objectives, where the latter encompasses a range of welfare functions, including utilitarian social welfare, Nash social welfare, and egalitarian welfare. Our algorithm achieves an 8n-approximation ratio for the Nash social welfare objective, which is a significant improvement over the previously known approximation ratio of O(n . log n). Moreover, for any given p, our algorithm computes an allocation with p-mean welfare at least 1/8n times the optimal. Our results hold for the wide range of subadditive valuations, including XOS and submodular valuations. We also show that our approximation guarantees are essentially tight for XOS valuations. Maximizing Nash social welfare for fair coverage We present a polynomial-time algorithm for maximizing Nash social welfare in coverage problems. We consider the problem of selecting T subsets of agents that achieve fair and efficient coverage while satisfying combinatorial constraints. We propose a valuation function based on the number of subsets that contain each agent, and design an algorithm that achieves an (18 + o(1))-approximation ratio for maximizing Nash social welfare in coverage instances. Our algorithm applies to instances where an FPTAS for weight maximization exists, and we complement our algorithmic result by proving that Nash social welfare maximization is APX-hard in coverage instances. Part-II: Individual Fairness Fair division using subsidy under dichotomous valuations We provide a subsidy-based algorithm for achieving envy-freeness in the allocation of indivisible goods among agents with dichotomous valuations. We show that it is possible to allocate goods among agents with dichotomous valuations in an envy-free manner with a per-agent subsidy of either 0 or 1, and such an envy-free solution can be computed efficiently in the standard value-oracle model. Our results hold for general dichotomous valuations, including non-additive and non-submodular valuations, and our subsidy bounds are tight, providing a linear improvement over the bounds known for general monotone valuations. DTSTART:20230519T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230518T120000Z UID:28665d9359e29b8db8f918f484c8d4ed-464 DTSTAMP:19700101T120011Z DESCRIPTION:Batch Proofs are Statistically Hiding URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/464/batch-proofs-are-statistically-hiding/ SUMMARY:Batch proofs are (possibly interactive) proof-systems that convince a verifier that x_1,...,x_t are in L, for some NP language L, with communication that is much shorter than sending the t witnesses. Batch proofs have been studied in various settings in recent years. In the case of statistical soundness (where the cheating prover is unbounded but honest prover is efficient), interactive solutions are known for any UP language. In the case of computational soundness (aka arguments, where both honest and dishonest provers are efficient), non-interactive solutions are now known for all of NP, assuming standard cryptographic assumptions. We study the necessary conditions for the existence of batch proofs in these two settings. Our main results are as follows. 1. Statistical Soundness: the existence of a statistically-sound batch proof for L implies that L has a Statistically Witness Indistinguishable (SWI) proof, with inverse polynomial soundness and SWI errors, and a non-uniform honest prover. In particular, under the assumption that NP does not have such SWI proofs, then batch proofs for all of NP do not exist. 2. Computational Soundness: the existence of batch arguments (BARGs) for NP, together with one-way functions, implies the existence of statistical zero-knowledge (SZK) arguments for NP with roughly the same number of rounds, and an inverse polynomial zero-knowledge error and non-uniform honest prover. Thus, constructing constant-round interactive BARGs from one-way functions would yield constant-round SZK arguments, which are currently only known assuming constant-round statistically-hiding commitments. 3. Non-interactive: the existence of non-interactive BARGs for NP, satisfying a notion of soundness called ``somewhere extractability`` (achieved by recent constructions), imply non-interactive statistical zero-knowledge arguments (NISZKA) for NP, with negligible zero-knowledge and soundness errors and a non-uniform honest prover. All of our results stem from a common framework showing how to transform a batch protocol for a language L into an SWI protocol for a single instance of L. (Based on work with Nir Bitansky, Omer Paneth, Prashant Nalini Vasudevan and Ron D. Rothblum) DTSTART:20230518T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230519T120000Z UID:f663f89fb5032aed9d891bebbf297988-465 DTSTAMP:19700101T120015Z DESCRIPTION:Exploring Welfare Maximization and Fairness in Participatory Budgeting URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/465/exploring-welfare-maximization-and-fairness-in-participatory-budgeting/ SUMMARY:Participatory budgeting (PB) is a voting paradigm for distributing a divisible resource, usually called a budget, among a set of projects by aggregating the preferences of individuals over these projects. It is implemented quite extensively for purposes such as government allocation of funds to public projects and funding agencies selecting research proposals to support. This dissertation aims to define welfare related and fairness related objectives for different PB models and proposes novel PB mechanisms that optimize these objectives. The thesis is divided into two main parts, focusing on dichotomous and ordinal preferences, respectively. Further, each part considers two cases: (i) the cost of each project is restricted to a single value and partial funding is disallowed and (ii) the cost of each project is flexible and can assume multiple values. Part 1: Participatory Budgeting under Dichotomous Preferences Welfare Maximization and Fairness under Restricted Costs Egalitarianism holds significance in PB, serving as both a welfare and fairness objective. Our work introduces and studies a natural egalitarian rule, Maxmin Participatory Budgeting (MPB). The study consists of two parts: computational analysis and axiomatic analysis. In the computational part, we prove that MPB is weakly NP-hard and propose FPT algorithms and an approximation algorithm. We also establish an upper bound on the achievable approximation ratio for exhaustive strategy-proof PB algorithms. In the axiomatic part, we investigate MPB by generalizing existing axioms and introducing a new fairness axiom called maximal coverage, which MPB satisfies. Welfare Maximization under Flexible Costs In this work, we present a model where projects have a discrete set of permissible costs, reflecting different levels of project sophistication. Voters express their preferences by specifying upper and lower cost bounds for each project. The outcome of a PB rule involves selecting a subset of projects and determining their costs. We explore different utility concepts and welfare-maximizing rules. We demonstrate that positive findings from single-cost projects can be extended to our framework with multiple permissible costs and we further analyze the fixed parameter tractability of the problems. We also propose axioms rich on intuition and evaluate their compatibility with PB rules. Part 2: Participatory Budgeting under Ordinal Preferences Welfare Maximization and Fairness under Restricted Costs This work focuses on incomplete weakly ordinal preferences and has two logical components. In the first component, we introduce dichotomous translation rules and the PB-CC rule which respectively expand on existing welfare-maximizing rules for dichotomous and strictly ordinal preferences. We show that our expansions largely maintain and even enhance the computational and axiomatic properties of these rules. We also propose a new relevant axiom, pro-affordability. The second component introduces the novel class of average rank-share guarantee rules to address fairness in participatory budgeting with ordinal preferences, overcoming limitations of existing fairness concepts in the literature. Fairness under Flexible Costs This work assumes that project costs have no restrictions, thereby making the model equivalent to random social choice. We investigate fairness in social choice under single-peaked preferences. Existing literature has extensively examined the construction and characterization of social choice rules in the single-peaked domain. We extend these findings by incorporating fairness considerations. To address group-fairness, we partition voters into logical groups based on attributes like gender or location. We introduce group-wise anonymity to capture fairness within each group and propose weak and strong notions of fairness to ensure fairness across groups. We characterize deterministic and random social choice rules that achieve group-fairness. We also explore the case without groups and provide more precise characterizations of rules achieving individual-fairness. DTSTART:20230519T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230519T120000Z UID:51af098cbf59cb8b3f5375a138b45c71-467 DTSTAMP:19700101T120012Z DESCRIPTION:A Productive and Scalable Actor-Based Programming System for PGAS Applications URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/467/a-productive-and-scalable-actor-based-programming-system-for-pgas-applications/ SUMMARY:The Partitioned Global Address Space (PGAS) model is well suited for executing irregular applications on distributed HPC systems due to its efficient support for short, one-sided messages. In this talk, we introduce a new programming system for PGAS applications, in which point-to-point remote operations can be expressed as fine-grained asynchronous active messages. A key observation is that these applications can benefit significantly from an actor-based model that moves computations to data as opposed to the traditional HPC approach of moving data to computations. Our approach can also be viewed as extending the classical Bulk Synchronous Processing (BSP) model to a Fine-grained-Asynchronous Bulk-Synchronous Parallelism (FA-BSP) model. We will discuss the programming models and runtime systems developed in the Habanero Extreme Scale Software Research Laboratory to realize the FA-BSP execution model, and present recent results illustrating the benefits of this approach on current HPC systems. Looking to the future, we will discuss some initial work-in-progress for hardware support of the FA-BSP execution model being undertaken in the Flow-Optimized Reconfigurable Zones of Acceleration (FORZA) project led by Georgia Tech that is supported by the IARPA AGILE program. The FORZA project is pursuing a software-hardware co-design approach to address the signicant disruptions currently under way in HPC hardware and software. In hardware, there is a Pandoras box of new architectural approaches being proposed to sustain performance improvements beyond the end of Moore’s Law. In software, there is an increased urgency for enabling large-scale data analytics applications for societal benefits. To address these challenges, the FORZA project is focusing on large-scale graph analytics as an important exemplar of the challenges being faced by many PGAS applications that unfortunately perform below 1% efficiency on todays systems. We would like to acknowledge all members of the Habanero lab and all participants in the FORZA project from Georgia Tech, Cornelis Networks, Lucata, Tactical Computing Labs, UC Santa Barbara, and U. Notre Dame. The opinions in this talk are solely those of the speaker and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of any of these organizations, the ODNI, IARPA, or U.S. Government. BACKGROUND REFERENCES: 1. DTSTART:20230519T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230602T120000Z UID:05bfbb8e0c180644a4d884aea5e845e7-468 DTSTAMP:19700101T120016Z DESCRIPTION:Algorithms Approaching the Threshold for Semi-random Planted Clique URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/468/algorithms-approaching-the-threshold-for-semi-random-planted-clique/ SUMMARY:We design new polynomial-time algorithms for recovering planted cliques in the semi-random graph model introduced by Feige and Kilian. The previous best algorithms for this model succeed if the planted clique has size at least n2/3 in a graph with n vertices. Our algorithms work for planted-clique sizes approaching n1/2 — the information-theoretic threshold in the semi-random model and a conjectured computational threshold even in the easier fully-random model. This result comes close to resolving open questions by Feige and Steinhardt. <br> Joint work with Pravesh Kothari and David Steurer. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> <br> <br> We are grateful to the Kirani family for generously supporting the theory seminar series <br> <br> <br> Hosts: Rameesh Paul, Rahul Madhavan, Aditya Subramanian and Aditya Abhay Lonkar DTSTART:20230602T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230608T120000Z UID:53f7c773d9311c84025bf478063ac90e-469 DTSTAMP:19700101T120011Z DESCRIPTION:Towards Robustness of Neural Legal Judgment System URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/469/towards-robustness-of-neural-legal-judgment-system/ SUMMARY:Legal Judgment Prediction (LJP) implements Natural Language Processing (NLP) techniques to predict judgment results based on fact description. It can play a vital role as a legal assistant and benefits legal practitioners and regular citizens. Recently, the rapid advances of transformer-based pretrained language models led to considerable improvement in this area. However, empirical results show that existing LJP systems are not robust to adversaries and noise. Also, they cannot handle large-length legal documents. In this work, we explore the robustness and efficiency of LJP systems even in a low data regime. In the first part, we empirically verified that existing state-of-the-art LJP systems are not robust. We further provide our novel architecture for LJP tasks which can handle extensive text lengths and adversarial examples. Our model performs better than state-of-the-art models, even in the presence of adversarial examples of the legal domain. In the second part, we investigate the approach for the LJP system in a low data regime. We provide a novel architecture using a few-shot approach that is also robust to adversaries. We conducted extensive experiments on American, European, and Indian legal datasets in the few-shot scenario. Our model, though trained using the few-shot approach, performs as well as state-of-the-art models which are trained using large datasets in the legal domain. DTSTART:20230608T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230616T120000Z UID:3f0ed7476cb62c4493eb21e229e953fd-470 DTSTAMP:19700101T120011Z DESCRIPTION:Learning linear thresholds from label proportions URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/470/learning-linear-thresholds-from-label-proportions/ SUMMARY:In many ML applications, training labels are not available for individual feature-vectors. Instead, the feature-vectors are grouped into sets or bags and we are given only the sum or the average of the labels in each bag. This is called Learning from Label Proportions (LLP) which is a direct generalization of traditional supervised learning. In this talk we shall motivate and formally define the LLP problem and provide the first study of computational learnability in the LLP setting, specifically of linear threshold functions (halfspaces). It is easy to see that linear programming no longer works as in the fully supervised setting. Indeed, we show that even in the realizable case with bags of size at most q, it is NP-hard to find a halfspace satisfying more than (1/q + o(1))-fraction of the bags. In particular, the problem is intractable even with bags of size at most 2. On the algorithmic side, we give a semi-definite programming based 2/5 approximation for bags of size 2. We also provide a 1/12 approximation for bags of size at most 3 -- using a more involved SDP and novel linear algebraic tools -- along with weaker guarantees for larger bags. <br> Based on two papers by the speaker which appeared in NeurIPS21 and NeurIPS22. <br> <br> Microsoft Teams link: <br> <a href="https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d">https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d</a> DTSTART:20230616T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230622T120000Z UID:1bfbc881838c7c4fecfae0d33085fa1a-471 DTSTAMP:19700101T120016Z DESCRIPTION:Multi-robot planning strategies for spatio-temporal sampling URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/471/multi-robot-planning-strategies-for-spatio-temporal-sampling/ SUMMARY:Understanding environmental and ecological processes, such as the health of coral reefs, or plankton blooms, or oil spills, requires persistent sampling of physical phenomena at spatio-temporal scales. In my research, I propose that efficient robotic sampling of a spatial field with regions of high local variance (hotspots) requires adaptive non-uniform sampling techniques. In this talk, I present algorithms and strategies to achieve efficient robotic sampling and reconstruction of non-uniform spatial fields. Spatial fields commonly occurring in nature consist of hotspots exhibiting extreme measurements and much higher spatial variability than the rest of the field, which is characterized by continuous, positively skewed, spatially correlated measurements. To collect data for modeling such fields, I apply informed path planning: a data collection strategy that computes paths to be traversed by a robot while considering resource constraints, such as power availability, and the uncertainty of the resulting model that should be minimized. I present informed non-myopic path planning techniques for robotic platforms to efficiently collect measurements from a spatio-temporal field and build a model of the underlying physical phenomenon with high accuracy. These methods are highly adaptive and run in real-time, which are key features for the methods to run on a real robot operating in real-world challenging environments. Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rachana Gusain, Rahul Madhavan, Rameesh Paul, KVN Sreenivas DTSTART:20230622T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230623T120000Z UID:7e5e1e2c7f363fe9b9702bc9afc7b698-472 DTSTAMP:19700101T120011Z DESCRIPTION:The Subspace Flatness Conjecture and Faster Integer Programming URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/472/the-subspace-flatness-conjecture-and-faster-integer-programming/ SUMMARY:In a seminal paper, Kannan and Lovasz (1988) considered a quantity μKL(Λ , K) which denotes the best volume-based lower bound on the covering radius μ(Λ , K) of a convex body K with respect to a lattice Λ . Kannan and Lovasz proved that μ(Λ , K) ≤ n. μKL(Λ , K) and the Subspace Flatness Conjecture by Dadush (2012) claims a O(log n) factor suffices, which would match the lower bound from the work of Kannan and Lovasz. We settle this conjecture up to a constant in the exponent by proving that μ(Λ , K) ≤ O(log3(n)).μKL(Λ , K). Our proof is based on the Reverse Minkowski Theorem due to Regev and Stephens-Davidowitz (2017). Following the work of Dadush (2012, 2019), we obtain a (log n)4n-time randomized algorithm to solve integer programs in n variables. Another implication of our main result is a near-optimal flatness constant of O (n log4(n)). Joint work with Thomas Rothvoss. Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rachana Gusain, Rahul Madhavan, Rameesh Paul, KVN Sreenivas DTSTART:20230623T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230623T120000Z UID:1eae1e83ccd26c4d73bedc48731f72a3-473 DTSTAMP:19700101T120016Z DESCRIPTION:Novel Algorithms for Improving Agricultural Planning and Operations using Artificial Intelligence and Game Theory URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/473/novel-algorithms-for-improving-agricultural-planning-and-operations-using-artificial-intelligence-and-game-theory/ SUMMARY:Farmer distress in developing and underdeveloped countries is a very common phenomenon. In-spite of several efforts put in by the various governments, NGOs, private companies, and the farmers themselves, alleviation of this problem remains elusive. Some of the major reasons behind the low returns, and losses, faced by the farmers are the inherent uncertainty in agriculture, unaffordability of advanced technologies, lack of access to markets, etc. This dissertation formulates and solves some of these contemporary problems in agriculture using artificial intelligence and game theory techniques. . Novel solutions are proposed that assist the state administration and the farmers during various stages of the agricultural crop cycle, starting from the pre-sowing and sowing decisions and going right up to the marketing of the produce. These solutions are: PREPARE (Price Prediction for Agriculture), ACRE (Agricultural Crop Recommendation Engine), CROP-S (Crop Planning System), and PROMISE (Procurement Mechanisms for Agricultural Inputs and Services). PREPARE (Price Prediction for Agriculture) Accurate prediction of agricultural crop prices is a crucial input for decision-making by various stakeholders in agriculture: farmers, consumers, retailers, wholesalers, and the Government. PREPARE accurately predicts crop prices using historical price information, climatic conditions, soil type, location, and other key determinants. In this direction, an innovative deep learning based approach is proposed, which achieves increased accuracy in price prediction. The proposed approach uses graph neural networks (GNNs) in conjunction with a standard convolutional neural network (CNN) model to exploit geospatial dependencies in prices. PREPARE works well with noisy legacy data and produces a performance that is at least 20% better than the results available in the literature. Accurate price prediction using PREPARE will significantly enhance the farmers and state administrations decision-making abilities. ACRE (Agricultural Crop Recommendation Engine) A key challenge faced by small and marginal farmers is to determine which crops to grow to maximize their utility. With a wrong choice of crops, farmers could end up with sub-optimal yields and low, and possibly even loss of revenue. ACRE (Agricultural Crop Recommendation Engine) is a tool that provides a scientific method to choose a crop or a portfolio of crops, to maximize the utility to the farmer. ACRE uses available data such as soil characteristics, weather conditions, and historical yield data, and uses machine learning/deep learning models to compute an estimated utility to the farmer. The main idea of ACRE is to generate several recommendations of portfolios of crops, with a ranking of portfolios based on the Sharpe ratio, a popular risk metric in financial investments. ACRE provides a rigorous, data-driven back-end for designing farmer-friendly mobile apps for assisting farmers in choosing crops. CROP-S (Crop Planning System) A mismatch between the crops produced by farmers and the respective market demands could potentially lead to large-scale crop dumping. This leads to huge financial losses and distress to the farmers. To alleviate this problem, CROP-S addresses the macro-level problem of district level or county level agricultural crop planning in any given state. Using CROP-S, the Government or any state administration can make an informed recommendation on which crop acreages (number of acres cultivated under each crop) to allocate in which districts or geospatial regions in a given state, so as to match the demand for the crops and maximize the profits for the farmers. CROP-S uses a the mathematical programming approach for maximizing the profits of farmers taking into account key determinants of farmer profits. CROP-S uses data about predicted demands, transportation costs, compliance ratios (fraction of farmers who will follow the recommended crop plan), and historical data about yields and prices to arrive at an optimal allocation of crop acreages to districts. PROMISE (Procurement Mechanisms for Agricultural Inputs and Services) Procuring agricultural inputs such as seeds, fertilizers, and pesticides, at desired quality levels and at affordable cost, forms a critical component of agricultural input operations. Farmer collectives (FCs), which are cooperative societies of farmers, offer an excellent prospect for enabling cost-effective procurement of inputs with assured quality to the farmers. The objective of PROMISE is to design sound, explainable mechanisms by which an FC will be able to procure agricultural inputs in bulk and distribute the inputs procured to the individual farmers who are members of the FC. In the methodology proposed, an FC engages qualified suppliers in a competitive, volume discount procurement auction in which the suppliers specify price discounts based on volumes supplied. The desiderata of properties for such an auction include: minimization of the total cost of procurement; incentive compatibility; individual rationality; fairness; and given business constraints. An auction satisfying all these properties is analytically infeasible. PROMISE uses a novel deep learning based approach to design such an auction. DTSTART:20230623T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230627T120000Z UID:4fb43b4b5db733dd85da65568efba547-474 DTSTAMP:19700101T120016Z DESCRIPTION:Data-driven Social Informatics URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/474/data-driven-social-informatics/ SUMMARY:Data-driven social informatics unites models derived from social science with data-driven approaches in order to model and predict population behavior patterns. It can be used to advance our understanding of human behavior, guide public policy decisions, and improve user experience with social media platforms. In this talk, I will describe work done at UCFs Intelligent Agents Lab (http://ial.eecs.ucf.edu/) in which we use a combination of agent-based modeling, natural language processing, and machine-learning to model human social systems for the DARPA challenge, ASIST (AI for Successful Teams). Our aim is to create multi-scale models for large-scale social media platforms, medium scale software engineering teams, and small scale human-agent interactions. Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rachana Gusain, Rahul Madhavan, Rameesh Paul, KVN Sreenivas DTSTART:20230627T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230628T120000Z UID:def23cdb31b86921ece13e7de132d192-475 DTSTAMP:19700101T120011Z DESCRIPTION:Algorithms for Achieving Fairness and Efficiency in Matching Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/475/algorithms-for-achieving-fairness-and-efficiency-in-matching-problems/ SUMMARY:Matching problems arise in numerous practical settings. Fairness and efficiency are two desirable properties in most such real-world settings. This dissertation work presents new approaches and algorithms to identify fair and/or efficient matchings. The thesis is organised into two logical parts: two sided preferences and single sided preferences. Part 1: Two Sided Preferences Incentive Compatibility in Stable Fractional Matchings We investigate the existence of incentive compatible mechanisms that find stable fractional matchings. We show, for general settings, that no incentive compatible mechanism can be stable. We characterise the space of instances that have a unique stable fractional matching. We prove for this set of instances that any stable matching mechanism will be incentive compatible. Fairness and Stability in Many-to-One Matchings We seek to optimize a fairness measure over the space of stable many-to-one matchings, motivated by a college admissions setting. With leximin optimality as the fairness notion, we first show the intractability of this problem. We identify a reasonable set of assumptions that makes this problem solvable in polynomial time. This requires that the agents on either side have the same ordinal rankings over the agents on the other side and that these preferences are strict. We show that on relaxing to weak rankings, the problem becomes APX-Hard. When we remove the ranking assumption but maintain strict preferences, the problem is NP-Hard. We show that the leximin optimal stable matching can be efficiently computed in the special case of two colleges. Part 2: Single Sided Preferences Repeated Matchings We propose a novel repeated matching model where the valuations of agents may change with how often they have received an item in the past. We study achieving fairness and efficiency separately as well as jointly in this setting. We find that optimizing for social welfare is NP-Hard for general valuations and tractable when the valuations are monotone with time. We also prove that maximizing for social welfare over the space of EF1 repeated matchings is NP-Hard. Further, we provide algorithms and non-existence results for EF1 and EFX repeated matchings in different settings. Fair and Efficient Delivery Motivated by the classical delivery problem, we introduce a novel model of fair division where delivery tasks must be fairly distributed among a set of agents. The delivery tasks are placed on the vertices of a given acyclic graph. The cost incurred by the agents is determined by the distance they travel from the hub where they start to service their assigned tasks. We study the existence of fair and efficient allocations of tasks to agents. We choose the fairness notions: EF1 and MMS, and efficiency notions: Pareto optimality and Social optimality. We find that while all these notions can be satisfied independently, the only combination of fairness and efficiency that can always be guaranteed is MMS and PO. For the remaining combinations, we provide characterisations of the space of instances for which they can be achieved. We find that most of the relevant problems are NP-Hard. We provide an XP-algorithm which finds the different combinations of fairness and efficiency whenever they exist. DTSTART:20230628T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230630T120000Z UID:5e51441c2fea4991b59d1fef9650902a-476 DTSTAMP:19700101T120016Z DESCRIPTION:Integrality Gaps for Random Integer Programs via Discrepancy URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/476/integrality-gaps-for-random-integer-programs-via-discrepancy/ SUMMARY:We prove new bounds on the additive gap between the value of a random integer program max cTx, Ax≤b, x∈{0,1}n with m constraints and that of its linear programming relaxation for a wide range of distributions on (A,b,c) . We are motivated by the work of Dey, Dubey, and Molinaro (SODA 21), who gave a framework for relating the size of Branch-and-Bound (B&B) trees to additive integrality gaps. Dyer and Frieze (MOR 89) and Borst et al. (Mathematical Programming 22), respectively, showed that for certain random packing and Gaussian IPs, where the entries of A,c are independently distributed according to either the uniform distribution on [0,1] or the Gaussian distribution N(0,1), the integrality gap is bounded by Om(log2n/n) with probability at least 1−1/n−e-Ω(m). In this paper, we generalize these results to the case where the entries of A are uniformly distributed on an integer interval (e.g., entries in {−1,0,1}), and where the columns of A are distributed according to an isotropic logconcave distribution. Second, we substantially improve the success probability to 1−1/poly(n), compared to constant probability in prior works (depending on m). Leveraging the connection to Branch-and-Bound, our gap results imply that for these IPs B&B trees have size npoly(m) with high probability (i.e., polynomial for fixed m), which significantly extends the class of IPs for which B&B is known to be polynomial. Our main technical contribution is a new linear discrepancy theorem for random matrices. Our theorem gives general conditions under which a target vector is equal to or very close to a {0,1} combination of the columns of a random matrix A. The proof uses a Fourier analytic approach, building on work of Hoberg and Rothvoss (SODA 19) and Franks and Saks (RSA 20). Joint work with Sander Borst (CWI) and Dan Mikunlincer (MIT). Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rachana Gusain, Rahul Madhavan, Rameesh Paul, KVN Sreenivas DTSTART:20230630T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230703T120000Z UID:310739e2973d0032536a15a013d406b5-477 DTSTAMP:19700101T120011Z DESCRIPTION:Round-or-Cut Technique for designing Approximation Algorithms for Clustering Problems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/477/round-or-cut-technique-for-designing-approximation-algorithms-for-clustering-problems/ SUMMARY:Many clustering problems are NP-hard and therefore extensive research has gone into designing approximation algorithms for these problems. Indeed, many techniques in approximation algorithms have been honed in the study of many of these fundamental problems. In this talk, I will talk about a technique called the “round-or-cut technique” (also called the “cutting plane technique” in integer programming parlance) which is probably not as well-known as many other techniques in approximation algorithms. In the last five years, however, this technique has led to tractable approximation algorithms for many clustering problems. I would like to present this general technique and illustrate it on one such clustering problem. This talk plans to be self-contained and no assumption of rounding or cutting will be assumed. Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rachana Gusain, Rahul Madhavan, Rameesh Paul, KVN Sreenivas DTSTART:20230703T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230704T120000Z UID:9b3dac677e3fa5de41eaeb0c71065cec-478 DTSTAMP:19700101T120015Z DESCRIPTION:SafeBet: A Simple, Secure and Fast Solution for Spectre and Meltdown URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/478/safebet-a-simple-secure-and-fast-solution-for-spectre-and-meltdown/ SUMMARY:Spectre and Meltdown attacks exploit microprocessor speculative execution to read and transmit forbidden data outside the attacker's trust domain and sandbox. Recent hardware schemes allow potentially-unsafe speculative accesses but prevent the secret's transmission by delaying all or many of the access-dependent instructions, even in the predominantly-common, no-attack case, which incurs performance loss and hardware complexity. Instead, we propose SafeBet which allows only, and in the common case does not delay most, safe accesses. We make the key observation that speculatively accessing a location is safe if the location has been accessed previously non-speculatively by the same trust domain (i.e., the location is within the domain's sandbox); and potentially unsafe, otherwise. We call the location as destination and the code memory region of the trust domain as the source. SafeBet employs the Speculative Memory Access Control Table (SMACT) to track non-speculative source address-destination address pairs. Disallowed accesses wait until reaching commit to trigger well-known replay without any intrusive hardware changes. SafeBet prevents all variants of Spectre and Meltdown except Lazy-FP-restore, based on any current or future side channel while using only simple, table-based access control and cache miss replay with virtually no change to the pipeline. Software simulations show that SafeBet uses 8.3 KB per core for the tables to perform within 6% on average (63% at worst) of the unsafe baseline behind which NDA-restrictive, a previous scheme of security and hardware complexity comparable to SafeBet's, lags by 83% on average. This work has been done in collaboration with Prof. T. N. Vijaykumar and our graduate students Conor Green and Cole Nelson. DTSTART:20230704T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230714T120000Z UID:c95247ec91062b3c306d5b4435bd9cac-479 DTSTAMP:19700101T120012Z DESCRIPTION:Algorithmic Problems on Vertex Deletion and Graph Coloring URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/479/algorithmic-problems-on-vertex-deletion-and-graph-coloring/ SUMMARY:In the thesis, we mainly discuss variants of two well-studied graph theoretical problems - vertex deletion problem and graph coloring problem - both have been used to tackle many real-world problems in diverse fields. Vertex deletion problems form a core topic in algorithmic graph theory with many applications. Typically, the objective of a vertex deletion problem is to delete the minimum number of vertices so that the remaining graph satisfies some property. Many classic optimization problems like Maximum Clique, Maximum Independent Set, and Vertex cover are examples of vertex deletion problems. In the first part of the thesis, we study popular vertex deletion problems called Cluster Vertex Deletion and its generalization s-Club Cluster Vertex Deletion, both being important in the context of graph-based data clustering. A cluster is often viewed as a dense subgraph (often a clique), and partitioning a graph into such clusters is one of the main objectives of graph-based data clustering. However, to account for the errors introduced during the construction of the network, the clusters of certain networks may be retrieved by making a small number of modifications such as deleting some vertices. Given a graph G, the objective of Cluster Vertex Deletion (CVD) is to delete a minimum number of vertices so that the remaining graph is a set of disjoint cliques. We focus on the polynomial-time solvability of CVD on special classes of graphs. Chordal graphs (graphs with no induced cycle of length greater than 3) are a well-studied class of graphs with many applications in algorithmic graph theory. Though polynomial-time algorithms for certain subclasses of chordal graphs such as interval graphs, block graphs, and split graphs are known, the computational complexity of CVD on chordal graphs remains unknown. We study CVD on well-partitioned chordal graphs, another subclass of chordal graphs that generalizes split graphs, which is introduced as a tool for narrowing down complexity gaps for problems that are hard on chordal graphs and easy on split graphs. In many applications, the equivalence of cluster and clique is too restrictive. For example, in protein networks where proteins are the vertices and the edges indicate the interaction between the proteins, a more appropriate notion of clusters may have a diameter of more than 1. Therefore researchers have defined the notion of s-clubs. An s-club is a graph with diameter at most s. The objective of s-Club Cluster Vertex Deletion (s-CVD) is to delete the minimum number of vertices from the input graph so that each connected component of the resultant graph is an s-club. We propose a polynomial-time algorithm for (s-CVD) on trapezoid graphs, a class of intersection graphs. To the best of our knowledge, our result provides the first polynomial-time algorithm for Cluster Vertex Deletion on trapezoid graphs. We also provide a faster algorithm for s-CVD on interval graphs. For each s >=1, we give an O(n(n+m))-time algorithm for s-CVD on interval graphs with n vertices and m edges. We also prove some hardness results for s-CVD on planar bipartite graphs, split graphs, and well-partitioned chordal graphs for each s >=2. Graph coloring has diverse applications and is still a prominent research area to tackle many practical problems by simulating them as coloring the vertices or edges of a graph subject to some constraints. In the second part of the thesis, we discuss a variant of the graph coloring problem. Efficient and scalable implementation of parallel algorithms on multiprocessor architectures with multiple memory banks requires simultaneous access to the data items. Such “conflict-free” access to parallel memory systems and other applied problems motivate the study of rainbow coloring of a graph, in which there is a fixed template T (or a family of templates), and one seeks to color the vertices of an input graph G with as few colors as possible, so that each copy of T in G is rainbow colored, i.e., has no two vertices the same color. We call such coloring a template-driven rainbow coloring and study the rainbow coloring of proper interval graphs (as hosts) for cycle templates. Generalizing the perfect graphs, Gyarfas introduced the concept of a Chi-bounded class of graphs. A graph class is said to be Chi-bounded if the chromatic number of the graphs in the class can be bounded by a function of their clique number. In the second part of the thesis, we also study the Chi-boundedness of squares of bipartite graphs and their subclasses. We also propose an approximation algorithm for the proper coloring of squares of convex bipartite graphs. DTSTART:20230714T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230704T120000Z UID:19e04736c593b7af2413c46982053fc1-480 DTSTAMP:19700101T120010Z DESCRIPTION:On Dynamics-Informed Blending of Machine Learning and Microeconomics URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/480/on-dynamics-informed-blending-of-machine-learning-and-microeconomics/ SUMMARY:Statistical decisions are often given meaning in the context of other decisions, particularly when there are scarce resources to be shared. Managing such sharing is one of the classical goals of microeconomics, and it is given new relevance in the modern setting of large, human-focused datasets, and in data-analytic contexts such as classifiers and recommendation systems. I will discuss several recent projects that aim to explore the interface between machine learning and microeconomics, including leader/follower dynamics in strategic classification, a Lyapunov theory for matching markets with transfers, and the use of contract theory as a way to design mechanisms that perform statistical inference. DTSTART:20230704T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230706T120000Z UID:b78a73b72d3cee1ad71a594029fe4115-481 DTSTAMP:19700101T120021Z DESCRIPTION:Unique Games and Expansion URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/481/unique-games-and-expansion/ SUMMARY:The Unique Games Conjecture (UGC) is a central open question in computational complexity and algorithms. In short, the UGC stipulates that almost satisfiable instances of Unique Games, a certain 2-variable constraint satisfaction problem (CSP), are NP-hard to distinguish from highly unsatisfiable instances. The UGC is known to imply a vast number of hardness-of-approximation results in combinatorial optimization. In this talk I will discuss our results that give algorithms for Unique Games (UG) on a large class of restricted instances: certifiable small-set expanders and graphs with certifiable global hypercontractivity. Our first algorithm solves UG instances whenever low-degree sum-of-squares (SoS) proofs certify good bounds on the small-set-expansion of the underlying constraint graph. A more complicated version of our algorithm succeeds even when the constraint graph is not a small-set expander as long as the structure of non-expanding small sets is “characterized” by a low-degree SoS proof, a property we call certified global hypercontractivity. The latter algorithm gives new insights into the breakthrough 2-2 Games hardness result, in particular, gives us some evidence as to what hard instances of UG look like. Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rameesh Paul, Rachana Gusain, Rahul Madhavan, KVN Sreenivas DTSTART:20230706T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230711T120000Z UID:b216ca5e1b421fc83f0892cf757796c2-482 DTSTAMP:19700101T120016Z DESCRIPTION:MMD-regularized Optimal Transport. URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/482/mmd-regularized-optimal-transport/ SUMMARY:Optimal transport (OT) induced Wasserstein distance is a popular tool for comparing probability measures. Kantorovichs formulation for OT aims to find an optimal plan for the transport of mass between the source and the target distributions that incurs the least expected cost of transportation. While classical OT strictly enforces the marginals of the transport plan to match the source and target, Unbalanced Optimal Transport (UOT) is employed when one wants to relax this constraint. In this talk, we will discuss our study of the UOT problem where the marginal constraints are enforced using an Integral Probability Metric (IPM), complementing the prior works on f-divergence regularized UOT. In particular, the talk will focus on MMD-regularized UOT (MMD-UOT), a special case of our formulation. The talk will cover some of our theoretical results, including the metricity properties, sample efficiency of the proposed metrics and consistency of our convex-program-based estimators. We will also discuss how MMD-UOT can be seen as an interpolant between the MMD & the Wasserstein metrics. Finally, the talk will present how these convex programs can be solved efficiently and showcase their utility in applications, including two sample tests, domain adaptation and single-cell RNA sequencing. DTSTART:20230711T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230717T120000Z UID:0da9de52ec95d643e79cc5213c9bdad5-483 DTSTAMP:19700101T120011Z DESCRIPTION:Online Bidding Algorithms for Return-on-Spend Constrained Advertisers URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/483/online-bidding-algorithms-for-return-on-spend-constrained-advertisers/ SUMMARY:Online advertising has recently grown into a highly competitive and complex multi-billion-dollar industry, with advertisers bidding for ad slots at large scales and high frequencies. This has resulted in a growing need for efficient DTSTART:20230717T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230717T120000Z UID:adcad7bdc632e6b75e6b03fc6768e8a0-484 DTSTAMP:19700101T120014Z DESCRIPTION:Moments, Random Walks, and Limits for Spectrum Approximation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/484/moments-random-walks-and-limits-for-spectrum-approximation/ SUMMARY:We study lower bounds for the problem of approximating a one dimensional distribution given (noisy) measurements of its moments. We show that there are distributions on [−1, 1] that cannot be approximated to accuracy ε in Wasserstein-1 distance even if we know all of their moments to multiplicative accuracy (1 ± 2−Ω(1/ε)); this result matches an upper bound of Kong and Valiant [Annals of Statistics, 2017]. To obtain our result, we provide a hard instance involving distributions induced by the eigenvalue spectra of carefully constructed graph adjacency matrices. Efficiently approximating such spectra in Wasserstein-1 distance is a well-studied algorithmic problem, and a recent result of Cohen-Steiner et al. [KDD 2018] gives a method based on accurately approximating spectral moments using 2O(1/ε) random walks initiated at uniformly random nodes in the graph. As a strengthening of our main result, we show that improving the dependence on 1/ε in this result would require a new algorithmic approach. Specifically, no algorithm can compute an ε-accurate approximation to the spectrum of a normalized graph adjacency matrix with constant probability, even when given the transcript of 2Ω(1/ε) random walks of length 2Ω(1/ε) started at random nodes. Based on joint work with Yujia Jin, Christopher Musco, and Aaron Sidford. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rameesh Paul, Rachana Gusain, Rahul Madhavan, KVN Sreenivas DTSTART:20230717T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230718T120000Z UID:32f6c981629eaf7d3e83cf72552f7d1d-485 DTSTAMP:19700101T120011Z DESCRIPTION:Tackling Label Corruptions: Univariate Polynomial Regression and Generalized Linear Models URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/485/tackling-label-corruptions-univariate-polynomial-regression-and-generalized-linear-models/ SUMMARY:Label corruptions pose a significant challenge in various machine learning tasks, affecting the accuracy and reliability of models. In this talk, we will address two distinct problems involving label corruptions, and present approaches to handle them effectively. In this problem the goal is to recover a polynomial widehat p which is pointwise close to a polynomial p, given samples (x, y) where, with probability1-rho the samples are clean, i.e. satisfy |y - p(x)| < sigma; and with probability rho is corrupted, i.e. completely arbitrary. We propose an approach which can tolerate rho as large as any constant less than 1/2, which is the information theoretic limit for unique recovery of this problem. In the second problem, we examine the challenge of learning a linear function composed with a generalized linear model, when upto 1-o(1) (i.e. all but a vanishingly small fraction) of the labels are corrupted via arbitrary independent and additive noise. We show that in this extremely challenging setting, it is always possible to recover a polynomial-sized list of candidates, one of which is arbitrarily close to the true answer. Moreover, if a certain mild identifiability condition holds, then it is possible to prune the list and return a single such candidate. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rameesh Paul, Rachana Gusain, Rahul Madhavan, KVN Sreenivas DTSTART:20230718T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230801T120000Z UID:62ae13195cfa67626fb144741b917330-486 DTSTAMP:19700101T120010Z DESCRIPTION:AVERAGE REWARD ACTOR-CRITIC WITH DETERMINISTIC POLICY SEARCH URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/486/average-reward-actor-critic-with-deterministic-policy-search/ SUMMARY:The average reward criterion is relatively less studied as most existing works in the Reinforcement Learning literature consider the discounted reward criterion. There are few recent works that present on-policy average reward actor-critic algorithms, but average reward off-policy actor-critic is relatively less explored. In this work, we present both on-policy and off-policy deterministic policy gradient theorems for the average reward performance criterion. Using these theorems, we also present an Average Reward Off-Policy Deep Deterministic Policy Gradient (ARO-DDPG) Algorithm. We first show asymptotic convergence analysis using the ODE-based method. Subsequently, we provide a finite time analysis of the resulting stochastic approximation scheme with linear function approximator and obtain an $epsilon$-optimal stationary policy with a sample complexity of $Omega(epsilon^{-2.5})$. We compare the average reward performance of our proposed ARO-DDPG algorithm and observe better empirical performance compared to state-of-the-art on-policy average reward actor-critic algorithms over MuJoCo-based environments. DTSTART:20230801T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230727T120000Z UID:6d5c4d0a77d787ef97e2c895c8f5841c-487 DTSTAMP:19700101T120011Z DESCRIPTION:Specification Synthesis with Constrained Horn Clauses URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/487/specification-synthesis-with-constrained-horn-clauses/ SUMMARY:Many practical problems in software development, verification and testing rely on specifications. The problem of specification synthesis is to automatically find relational constraints for undefined functions, called specifications, in a given program and a postcondition. Specifications are helpful in verifying open programs, compositional verification, automatic test case generation, run-time checks, program understanding and many more applications. Maximal (or logically weakest) specifications are desirable in all these applications. The specifications also have to be valid with respect to the given program's semantics and the postcondition.        Existing tools expect the developers of programs to provide specifications. Providing specifications is tedious for developers, hindering the wide adoption of program verification and synthesis tools. Automatic synthesis of valid and maximal specifications is beneficial; but has several challenges. This thesis proposes techniques to overcome the challenges.              This thesis proposes a framework called Infer-Check-Weaken, using which valid and maximal specifications are effectively synthesised. The core idea behind Infer-Check-Weaken is first to infer a valid specification. Then, in a loop, the specification is checked for maximality. When the check fails, weakening is performed. This loop continues till a maximal specification is obtained. Depending on the specifications to be inferred and the structure of the programs, we propose different techniques to perform each task. We show the effectiveness of these techniques over publicly available benchmarks.        More specifically, for programs with integer variables and linear arithmetic operations, we propose an inference technique called Non-vacuous Specification Synthesis (NSS), which can find specifications like function contracts on undefined functions, preconditions, and others along with inductive invariants. NSS works on the idea of backward and forward reasoning. The maximality checking is performed by encoding the problem as a logical formula and then using an SMT solver. Further, weakening is done using the feedback from maximality checking and the NSS module. This thesis's second and third contributions are finding the weakest precondition for deterministic and non-deterministic terminating programs that manipulate arrays. Such programs often require quantified preconditions, which are difficult to reason about. To infer quantified preconditions, we propose two techniques: Range Abduction (RA) and its extension Structural Array Abduction (SAA). RA extends the logical abduction to quantified formulas over arrays. SAA further extends RA to find proof of maximality additionally. We propose different maximality-checking techniques for deterministic and non-deterministic programs based on program transformation and a syntax-guided-synthesis-based weakening procedure. These techniques work at the level of verification conditions in the Floyd-Hoare style, making them agnostic to programming languages. These verification conditions are called constrained Horn clauses, which are gaining popularity as a logical intermediate representation among variousx program verification tools. Hence, our techniques can be integrated with the existing tools. DTSTART:20230727T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230721T120000Z UID:ca4e12885a32c1b5baa04d5223c4109d-489 DTSTAMP:19700101T120010Z DESCRIPTION:Temporal Point Processes for Forecasting Events in Higher-Order Networks URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/489/temporal-point-processes-for-forecasting-events-in-higher-order-networks/ SUMMARY:Complex systems consisting of interacting entities can be effectively represented as time-evolving networks or graphs, where the entities are depicted as nodes, and the interactions between them are represented as instantaneous edges. Modeling the evolution of these systems and forecasting interaction events are of significant importance for many fields, such as e-commerce, financial markets, neuroscience, etc. This is achieved using the Temporal Point Process (TPP) framework, a stochastic process that models these interactions as discrete events occurring in continuous time. The existing works on interaction forecasting are applicable only to pair-wise edges. However, real-world events involving interactions are much more complex than pair-wise interactions. It involves a group of entities interacting in a complex way rather than just two entities. This leads to the formation of time-evolving higher-order networks. There has not been much research to develop machine learning algorithms for event prediction in these types of networks. This thesis addresses this by providing solutions to the following problems: (i) How can we use TPP to forecast higher-order interaction events? (ii) Considering the number of possible events grows exponentially when problem setting changes from pair-wise to higher-order, can we forecast the next event in a scalable way? (iii) How can we incorporate relations and group structure within a higher-order interaction into the forecasting model? The first contribution of this thesis is a model for forecasting higher-order interactions among a group of entities as instantaneous hyperedge events in a network. In this model, we introduce a TPP on each hyperedge, with the conditional intensity parameterized by a hyperedge-based decoder that uses node embeddings. To account for the temporal evolution of entities, we employ Temporal Graph Representation Learning (TGRL) techniques such as hypergraph convolution to learn dynamic node embeddings, and model parameters are learned by minimizing the negative log-likelihood of the TPP process. Furthermore, our model has been extended to accommodate bipartite interactions, where interactions occur between two distinct groups of entities of different types. To achieve this, we introduce a bipartite hyperedge-based decoder, which incorporates separate node embedding modules for each node type. Through comprehensive experiments, we demonstrate the effectiveness of our model by comparing it to previous works and baseline models. Moreover, through ablation studies, we highlight the superior performance of hyperedge-based models in capturing higher-order interactions compared to pairwise models. Secondly, we introduce a model to forecast directed higher-order interactions occurring between two distinct groups of entities. Unlike the previous approach that focuses on representation learning from higher-order interactions, here we also introduce a strategy to forecast hyperedges in a scalable way. For that, we employ a multi-task framework for forecasting candidate hyperedges. This involves a TPP-based model to predict the time of events on each node, followed by pairwise neighborhood and hyperedge size prediction modules for generating candidate hyperedges. This will reduce the exponential search in forecasting future hyperedges in the previous models. Then a directed hyperedge link predictor is used to identify the true hyperedge from false ones. Further, we devise a TGRL framework that improves the models scalability when dealing with datasets containing many interactions using a memory network module. In our final contribution, we extend the existing higher-order interaction forecasting approaches based on TPP to accommodate real-world interactions that involve internal group structures with nodes of different types. Each group is associated with a specific relation type, and to address this complexity, we introduce the concept of multi-relational recursive hyperedge formation events. In this framework, hyperedges can serve as nodes within other hyperedges, creating a hierarchical structure. Additionally, we extended the widely-used Temporal Knowledge Graphs (TKG) framework within this context by incorporating subject-specific and object-specific hyperedges. In conclusion, this thesis emphasizes and demonstrates the importance of employing higher-order dynamic network models to forecast interactions in real-world complex systems effectively. As part of this, we have shown that hyperedges can represent higher-order interactions as they provide a natural framework for representing relations between more than two entities. Our first work uses it for higher-order interaction forecasting of homogeneous and bipartite nature. This is further extended to directed higher-order interactions in the second work, where we show the use of directed hyperedges. In this, we tackled the challenge of scalability in forecasting higher-order interactions by adopting a TPP-based multi-task approach. Through our final contribution, we showed the need to go beyond hyperedge by using multi-relation recursive hyperedges to incorporate the intricate relations of entities in a higher-order interaction. DTSTART:20230721T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230727T120000Z UID:fe478a4e02d4935bcaf9fa8556d7b311-491 DTSTAMP:19700101T120011Z DESCRIPTION:An MLIR-based High-level Synthesis Compiler for Hardware Accelerator Design URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/491/an-mlir-based-high-level-synthesis-compiler-for-hardware-accelerator-design/ SUMMARY:The emergence of machine learning, image and audio processing on edge devices has motivated research towards power-efficient custom hardware accelerators. Though FPGAs are an ideal target for custom accelerators, the difficulty of hardware design and the lack of vendor-agnostic, standardized hardware compilation infrastructure has hindered their adoption. High-level synthesis (HLS) offers a more compiler-centric alternative to the traditional Verilog-based hardware design improving developer productivity. In this work, we propose an MLIR-based end-to-end HLS compiler and an an intermediate representation that is suitable for the design and implementation of domain-specific accelerators for affine workloads. Our compiler brings similar levels of modularity and extensibility to the HLS compilation domain, which LLVM brought to the area of a software compilation. A modular compiler infrastructure offers the advantage of incrementally introducing new language frontends and optimization passes without the need to reinvent the whole HLS compiler stack. Our compiler converts a high-level description of the accelerator specified in the C programming language into a register-transfer-level design in SystemVerilog. We use memory dependence analysis and integer-linear-program(ILP) based automatic scheduling on improving loop-pipelining and introduce parallelization between producer-consumer kernels. Our ILP-based optimizer beats the state-of-the-art Vitis HLS compiler by 1.3x in performance over a representative set of benchmarks while requiring fewer FPGA resources. Microsoft teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_MmVlOTllMzMtMWU2ZC00ZmQ2LWFmMDYtNDRkMjgxMDE5OTI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%22171d9abc-cf43-429a-9680-c05b9523fa9a%22%7d DTSTART:20230727T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230804T120000Z UID:58239ac7375c6efc7df1838482d8a8c0-492 DTSTAMP:19700101T120015Z DESCRIPTION:Data-Aware Network-on-Chip for High End Computing Systems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/492/data-aware-network-on-chip-for-high-end-computing-systems/ SUMMARY:Tiled Chip Multi-Processors (TCMP) is one of variant of Multiprocessor System on Chips (MPSoCs) MPSoC that uses Network on Chip (NoC) for inter tile communication. NoC is a multi-hop packet-based communication infrastructure that connects different cores with each other through routers. NoC provides a high transfer bandwidth, and the infrastructure is scalable. Applications running on different cores access on-chip last level cache memories and off-chip main memory through the underlying NoC. As NoC plays a vital role in memory access latency, ignoring the nature of its infrastructure may severely impact the performance of the applications in TCMPs. This talk attempts to share some thoughts on a few architectural optimizations done to establish a dynamic cooperation between NoC and the memory hierarchy for improved performance. By gathering information about the data travelling from one level of memory to another, we exploit underutilized NoC resources to design techniques and propose optimizations that reduce memory access latency and improve overall system performance. DTSTART:20230804T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230808T120000Z UID:4b4156e869ad79ebf4e1d754d363d6b9-493 DTSTAMP:19700101T120011Z DESCRIPTION:Decentralized Information Flow Control for the Robot Operating System URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/493/decentralized-information-flow-control-for-the-robot-operating-system/ SUMMARY:The Robot Operating System (ROS) is a popular open-source middleware widely used in the robotics community. While ROS provides extensive support for robotic application development, it lacks certain fundamental security features, making ROS-based systems vulnerable to attacks that can compromise the application and user security. To address these challenges, ROS incorporates security plugins and libraries to protect against unauthorized access and ensure secure communication between ROS applications. However, these user-level security tools do not protect end-to-end information flow against operating system (OS)-level attacks. This research introduces FlowROS, a decentralized information flow control (DIFC) system for ROS. FlowROS empowers ROS applications with fine-grained control over their sensitive information, providing a programmable interface and supporting explicit label propagation for modified ROS applications. FlowROS also leverages implicit label propagation for backward compatibility with unmodified ROS applications while guaranteeing end-to-end information flow control, including secrecy and integrity requirements. The implementation of FlowROS includes a kernel-level enforcement engine based on Linux security modules (LSM) to intercept sensitive communications within the system. The contributions of this research include identifying the limitations of mandatory access control (MAC)-based policy frameworks in ROS, motivating the need for DIFC systems in robotics platforms, presenting FlowROS as a practical DIFC solution for ROS applications, addressing the inherent DIFC challenge in ROS, and demonstrating the robustness, security, and performance of FlowROS through case studies, evaluations, and practical policies. Overall, FlowROS enhances the security of ROS-based systems by providing ROS applications explicit control over the flow of their sensitive information, mitigating vulnerabilities, and protecting against accidental data disclosure. Meeting Link : This will be an in-person event in CSA-254. DTSTART:20230808T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230804T120000Z UID:a20dd4fc3de4ab8d3f78d06d8279275e-494 DTSTAMP:19700101T120015Z DESCRIPTION:Data-Aware Network-on-Chip for High End Computing Systems URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/494/data-aware-network-on-chip-for-high-end-computing-systems/ SUMMARY:Tiled Chip Multi-Processors (TCMP) is one of variant of Multiprocessor System on Chips (MPSoCs) MPSoC that uses Network on Chip (NoC) for inter tile communication. NoC is a multi-hop packet-based communication infrastructure that connects different cores with each other through routers. NoC provides a high transfer bandwidth, and the infrastructure is scalable. Applications running on different cores access on-chip last level cache memories and off-chip main memory through the underlying NoC. As NoC plays a vital role in memory access latency, ignoring the nature of its infrastructure may severely impact the performance of the applications in TCMPs. This talk attempts to share some thoughts on a few architectural optimizations done to establish a dynamic cooperation between NoC and the memory hierarchy for improved performance. By gathering information about the data travelling from one level of memory to another, we exploit underutilized NoC resources to design techniques and propose optimizations that reduce memory access latency and improve overall system performance. DTSTART:20230804T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230809T120000Z UID:c850d025949b571cb16c3b314e11da9c-496 DTSTAMP:19700101T120011Z DESCRIPTION:The Price of Equity with Binary Valuations and Few Agent Types URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/496/the-price-of-equity-with-binary-valuations-and-few-agent-types/ SUMMARY:In fair division problems, the notion of price of fairness measures the loss in welfare due to a fairness constraint. Prior work on the price of fairness has focused primarily on envy-freeness up to one good (EF1) as the fairness constraint, and on the utilitarian and egalitarian welfare measures. Our work instead focuses on the price of equitability up to one good (EQ1) (which we term price of equity) and considers the broad class of generalized p-mean welfare measures (which includes utilitarian, egalitarian, and Nash welfare as special cases). We derive fine-grained bounds on the price of equity in terms of the number of agent types (i.e., the maximum number of agents with distinct valuations), which allows us to identify scenarios where the existing bounds in terms of the number of agents are overly pessimistic. Our work focuses on the setting with binary additive valuations, and obtains upper and lower bounds on the price of equity for p-mean welfare for all p⩽1. For any fixed p, our bounds are tight up to constant factors. A useful insight of our work is to identify the structure of allocations that underlie the upper (respectively, the lower) bounds simultaneously for all p-mean welfare measures, thus providing a unified structural understanding of price of fairness in this setting. This structural understanding, in fact, extends to the more general class of binary submodular (or matroid rank) valuations. We also show that, unlike binary additive valuations, for binary submodular valuations the number of agent types does not provide bounds on the price of equity. This is a joint work with Umang Bhaskar, Neeldhara Misra and Rohit Vaish, and will be presented at SAGT 23. The full version of the paper can be accessed here - https://arxiv.org/abs/2307.06726. Microsoft Teams link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGE3NDg5NzktMWQ0Zi00MzFmLTg5OTgtMTMyYWM4MWQyYjI2%40thread.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%227c84465e-c38b-4d7a-9a9d-ff0dfa3638b3%22%7d We are grateful to the Kirani family for generously supporting the theory seminar series Hosts: Rameesh Paul, Rachana Gusain, Rahul Madhavan, KVN Sreenivas DTSTART:20230809T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230808T120000Z UID:5534d434ec76d3a1cb3c38a7d2c5bdad-497 DTSTAMP:19700101T120016Z DESCRIPTION:Cooperation Competition in Partition form games URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/497/cooperation-competition-in-partition-form-games/ SUMMARY:Game theory has been widely employed to examine scenarios where agents act in their own self-interest. Even though agents may have self-serving motives, collaboration between them can be facilitated if they stand to benefit from working together. This brings forth the notion of cooperation among agents, leading to a cooperative game. Such games can be analyzed using tools from cooperative game theory. In literature, cooperative games are primarily analyzed in characteristic form and the stability of the grand coalition, comprising all players, is evaluated. This involves determining the existence of an allocation vector that discourages agents from deviating either independently or collectively from the grand coalition. However, it is also possible for a subset of agents to collaborate leading to a disjoint collections of agents, commonly referred to as partition. This gives rise to a coalition formation game, where each coalition operates independently and competes with other coalitions, while agents within each coalition work together to maximize their coalitions welfare. In such situations, the welfare of a coalition may be influenced by both the members within it and the arrangement of players outside of it, leading to a partition form game. This talk will focus on such games, their solution concepts and their applications in online auctions and queuing systems. DTSTART:20230808T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230809T120000Z UID:1da5f7c25bbebceb80121d984e8ca023-498 DTSTAMP:19700101T120016Z DESCRIPTION:Cooperation Competition in Partition form games URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/498/cooperation-competition-in-partition-form-games/ SUMMARY:Game theory has been widely employed to examine scenarios where agents act in their own self-interest. Even though agents may have self-serving motives, collaboration between them can be facilitated if they stand to benefit from working together. This brings forth the notion of cooperation among agents, leading to a cooperative game. Such games can be analyzed using tools from cooperative game theory. In literature, cooperative games are primarily analyzed in characteristic form and the stability of the grand coalition, comprising all players, is evaluated. This involves determining the existence of an allocation vector that discourages agents from deviating either independently or collectively from the grand coalition. However, it is also possible for a subset of agents to collaborate leading to a disjoint collections of agents, commonly referred to as partition. This gives rise to a coalition formation game, where each coalition operates independently and competes with other coalitions, while agents within each coalition work together to maximize their coalitions welfare. In such situations, the welfare of a coalition may be influenced by both the members within it and the arrangement of players outside of it, leading to a partition form game. This talk will focus on such games, their solution concepts and their applications in online auctions and queuing systems. DTSTART:20230809T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230816T120000Z UID:28debfd32417a45c442a8a406571054b-499 DTSTAMP:19700101T120010Z DESCRIPTION:Maximum Independent Set of Rectangles - An Empirical Study URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/499/maximum-independent-set-of-rectangles-an-empirical-study/ SUMMARY:We study the Maximum Independent Set of Rectangles (MISR) problem. The problem involves a collection of n axis-parallel rectangles with weights. The objective is to select the subset of non-overlapping rectangles with the maximum total weight. The unweighted setting is a special case where the weight of each rectangle is 1. The problem has many practical applications, such as map labelling, data mining, and resource allocation. The focus of our work is to conduct empirical evaluations on various algorithmic and combinatorial questions related to the MISR problem. The thesis investigates the following problems and presents empirical results: 1. The current state-of-the-art in the MISR problem: Chalermsook et al. (2020) proposed an O(log log n)- factor approximation algorithm for the problem. In the unweighted setting, Galvez et al. (2021) proposed a 3-factor approximation algorithm. However, both algorithms are theoretical but not practical. We have implemented two practical algorithms, namely a linear programming (LP)-based algorithm and a greedy algorithm for the MISR problem. We have conducted experimental studies on these algorithms, using randomly generated and special distributions of axis-parallel rectangles. Based on our experiments, we observed that the LP-based approach obtained an independent set which is at most 1% away from the optimum, whereas the independent set generated by the greedy approach is at most 10% away from the optimum. 2. We have empirically evaluated three conjectures in the intersection graph of axis parallel rectangles involving the chromatic number (χ), clique number (ω), independence number (α), and piercing number (τ). These conjectures have been studied well and are considered challenging open problems in the area of combinatorial geometry. We have implemented simple algorithms to compute the quantities χ, ω, α, and τ. We have conducted experimental studies to evaluate these conjectures, using randomly generated and special distributions of axis-parallel rectangles. The first conjecture is: χ/ω = O(1). The best known upper bound is O(log ω), given by Chalermsook et al. (2020). In our experimental study, we have found that the ratio χ/ω is at most 1.38. The second conjecture is: τ /α = O(1). The best known upper bound for the ratio is O((log log α)2), given by Correa et al. (2014). In our experimental study, we have found that the ratio τ /α is at most 1.27. The third conjecture is: ω·α/n = Ω(1). The best known lower bound for the ratio is Ω(1/(log log α)2). In our experimental study, we have found that the ratio ω·α/n is at least 2.71. Thus, our empirical study validates these conjectures for both random and special distributions of axis-parallel rectangles. DTSTART:20230816T120000Z END:VEVENT BEGIN:VEVENT DTEND:20230810T120000Z UID:b4ecf28f46aea9589726545173a11242-500 DTSTAMP:19700101T120016Z DESCRIPTION:Robust fake-post detection against real-coloring adversaries: Branching process and Stochastic approximation URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/500/robust-fake-post-detection-against-real-coloring-adversariesbranching-process-and-stochastic-approximation/ SUMMARY:The viral propagation of fake posts on online social networks (OSNs) has become an alarming concern. We design control mechanisms for fake post detection while negligibly affecting the propagation of real posts. Towards this, a warning mechanism based on crowd-signals was recently proposed, where all users actively declare the post as real or fake. In the talk, we will discuss a more realistic framework where users exhibit different adversarial or non-cooperative behaviour: (i) they can independently decide whether to provide their response, (ii) they can choose not to consider the warning signal while providing the response, and (iii) they can be real-coloring adversaries who deliberately declare any post as real. To analyze the post-propagation process in this complex system, we propose and study a new branching process, namely total-current population-dependent branching process with multiple death types. For the branching process, under finite second-moment conditions, using stochastic approximation technique, we show that the time-asymptotic proportion of the populations either converges to the equilibrium points or infinitely often enters every neighbourhood and exits some neighbourhood of a saddle point of an appropriate ordinary differential equation with a certain probability. For the application at hand, at first, we compare and show that the existing warning mechanism significantly under-performs in the presence of adversaries. Then, we design new mechanisms which remarkably perform better than the existing mechanism by cleverly eliminating the influence of the responses of the adversaries. Towards the end, we propose an algorithm which works the best, without assuming any prior knowledge about user specific parameters. The theoretical results are validated using Monte-Carlo simulations. DTSTART:20230810T120000Z END:VEVENT END:VCALENDAR