AIML Workshop for Stanford Students at CSA, IISc
Supported by the Walmart Centre for Tech Excellence
13th September 2024 (Friday)
Time : 15:30 to 18:30
Location : CSA 104
The Department of Computer Science and Automation (CSA) at the Indian Institute of Science (IISc) is excited to host undergraduate students from Stanford University for a half-day workshop focused on cutting-edge research in Artificial Intelligence and Machine Learning (AIML).
Inauguration by CSA Chair:
Prof. Vinod Ganapathy
-
The GRAMA Project: Empowering Farmers with AI-Based Algorithms
Prof. Y. Narahari
An insight into how AI is transforming agricultural practices in rural India through the GRAMA project. -
Designing Salient, Naturalistic “Super-Stimuli” with Deep Generative Models
Prof. Sridharan Devarajan
A discussion on using deep generative models to create super-stimuli for neuroscience applications. -
Challenges in Embodied AI: What Works and What Doesn’t
Prof. Shishir Kolathaya
Explore the hurdles faced by AI systems interacting in the physical world and their potential solutions.
Poster Session by IISc Students
Engage with cutting-edge research presented by students from IISc.
Light Snacks and Networking
A casual networking session to wrap up the day.
Poster Presentation Details
Abstract : Reinforcement learning has traditionally been studied with exponential discounting or the average reward setup, mainly due to their mathematical tractability. However, such frameworks fall short of accurately capturing human behavior, which has a bias towards immediate gratification. Quasi-Hyperbolic (QH) discounting is a simple alternative for modeling this bias. Unlike in traditional discounting, though, the optimal QH-policy, starting from some time $t_1,$ can be different to the one starting from $t_2.$ Hence, the future self of an agent, if it is naive or impatient, can deviate from the policy that is optimal at the start, leading to sub-optimal overall returns. To prevent this behavior, an alternative is to work with a policy anchored in a Markov Perfect Equilibrium (MPE). In this work, we propose the first model-free algorithm for finding an MPE. Using a brief two-timescale analysis, we provide evidence that our algorithm converges to invariant sets of a suitable Differential Inclusion (DI). We also show that the QH Q-value function of any MPE would be an invariant set of our identified DI. Finally, we validate our claims numerically for the standard inventory system with stochastic demands. We believe our work significantly advances the practical application of reinforcement learning.
Abstract : “We study the probabilistic assignment of items to platforms that satisfies both group and individual fairness constraints. Each item belongs to specific groups and has a preference ordering over platforms. Each platform enforces group fairness by limiting the number of items per group that can be assigned to it. There could be multiple optimal solutions that satisfy the group fairness constraints, but this alone ignores item preferences. Our approach explores a best of both worlds fairness' solution to get a randomized matching, which is ex-ante individually fair and ex-post group-fair. Thus, we seek aprobabilistic individually fair’ distribution over group-fair' matchings where each item has ahigh’ probability of matching to one of its top choices. This distribution is also ex-ante group-fair. Users can customize fairness constraints to suit their requirements. Our first result is a polynomial-time algorithm that computes a distribution over `group-fair’ matchings such that the individual fairness constraints are approximately satisfied and the expected size of a matching is close to OPT. We empirically test this on real-world datasets. We present two additional polynomial-time bi-criteria approximation algorithms that users can choose from to balance group fairness and individual fairness trade-offs.
For disjoint groups, we provide an exact polynomial-time algorithm adaptable to additional lower group fairness' bounds. Extending our model, we encompassmaxmin group fairness,’ amplifying underrepresented groups, and `mindom group fairness,’ reducing the representation of dominant groups.'”
Abstract :Actor Critic methods have found immense applications on a wide range of Reinforcement Learning tasks especially when the state-action space is large. In this paper, we consider actor critic and natural actor critic algorithms with function approximation for constrained Markov decision processes (C-MDP) involving inequality constraints and carry out a non-asymptotic analysis for both of these algorithms in a non-i.i.d (Markovian) setting. We consider the long-run average cost criterion where both the objective and the constraint functions are suitable policy-dependent long-run averages of certain prescribed cost functions. We handle the inequality constraints using the Lagrange multiplier method. We prove that these algorithms are guaranteed to find a first-order stationary point (i.e., $\Vert \nabla L(\theta,\gamma)\Vert_2^2 \leq \epsilon$) of the performance (Lagrange) function $L(\theta,\gamma)$, with a sample complexity of $\mathcal{\tilde{O}}(\epsilon^{-2.5})$ in the case of both Constrained Actor Critic (C-AC) and Constrained Natural Actor Critic (C-NAC) algorithms. We also show the results of experiments on three different Safety-Gym environments.
Abstract : Structured pruning without access to the training data or the loss function is an active area of research. Key to this problem is identifying which filters are useful to the model in the constrained setting. Recently, distributional structured pruning in Deep Neural Networks was introduced to address this problem, which seeks to identify and retain discriminative filters that can effectively differentiate between classes. Crucial to such methods is the ability to estimate the discriminative ability of a filter. In this work, we define the discriminative ability as the error rate of the Bayes optimal classifier for the given distribution. Since this cannot be exactly computed in a general setting, we exploit the relationship between the Total Variation (TV) distance and the Bayes risk. However, computing the TV distance is also intractable, motivating us to we derive distribution-agnostic lower bounds on TV Distance that depend only on the moments of witness functions. Using this bound, we also establish new relationships between the TV Distance and well-known discriminant-based classifiers, such as Fisher Discriminants and Minimax Probability machines. The lower bounds are used to produce a variety of pruning algorithms called WitnessPrune by varying the choice of witness function. Our experiments show that our approach consistently matches or outperforms current state-of-the-art.
“The autoencoder is an unsupervised learning paradigm that aims to create a compact latent representation of data by minimizing the reconstruction loss. However, it tends to overlook the fact that most data (images) are embedded in a lower-dimensional latent space, which is crucial for effective data representation. To address this limitation, we propose a novel approach called Low-Rank Autoencoder (LoRAE). In LoRAE, we incorporated a low-rank regularizer to adaptively learn a low-dimensional latent space while preserving the basic objective of an autoencoder. This helps
embed the data in a lower-dimensional latent space while preserving important information. It is a simple autoencoder extension that learns low-rank latent space. Theoretically, we establish a tighter error bound for our model. Empirically, our model’s superiority shines through various
tasks such as image generation and downstream classification. Both theoretical and practical outcomes highlight the importance of acquiring low-dimensional embeddings.
“
Abstract : We establish the first finite-time logarithmic regret bounds for the self-tuning regulation problem. We introduce a modified version of the certainty equivalence algorithm, which we call PIECE, that clips inputs in addition to utilizing probing inputs for exploration. We show that it has a C log(T) upper bound on the regret after T time-steps for bounded noise, and C log^3(T) in the case of sub-Gaussian noise, unlike the LQ problem where logarithmic regret is shown to be not possible. The PIECE algorithm is also designed to address the critical challenge of poor initial transient performance of reinforcement learning algorithms for linear systems. Comparative simulation results illustrate the improved performance of PIECE.
Abstract : “Graph Neural Networks (GNNs) have experienced rapid advancements in recent years due to their ability to learn meaningful representations from graph data structures. Federated Learning (FL) has emerged as a viable machine learning approach for training a shared model on decentralized data, addressing privacy concerns while leveraging parallelism. Existing methods that address the unique requirements of federated GNN training using remote embeddings to enhance convergence accuracy are limited by their diminished performance due to large communication costs with a shared embedding server. In this paper, we present OpES, an optimized federated GNN training framework that uses remote neighbourhood pruning, and overlaps pushing of embeddings
to the server with local training to reduce the network costs and training time. The modest drop in per-round accuracy due to pre-emptive push of embeddings is out-stripped by the reduction in per-round training time for large and dense graphs like Reddit and Products, converging up
to ? 2× faster than the state-of-the-art technique using an embedding server and giving up to 20% better accuracy than vanilla federated GNN learning.”
Abstract : The frameworks for explaining the functional space learned by deep neural networks, also known as eXplainable AI (XAI) models, are majorly based on the notion of the locality. Most of the approaches for local model-agnostic explainability employ linear models. Driven by the fact that a linear model is inherently interpretable (linear coefficients being the explanation), they are used to approximate the non-linear function locally. In this paper, we argue that local linear approximation is inapt as the black boxes under investigation are often highly non linear. We present a novel perturbation-based approach for local explainability, called the Distillation Approach for Model-agnostic Explainability (DAME). It separates out the two tasks- local approximation and generating explanation, and successfully attempts generating explanations by operating on high dimensional input space. The DAME framework is a learnable, saliency-based explainability model, which is post-hoc, model-agnostic, and requires only query access to the black box. Extensive evaluations including quantitative, qualitative and subjective measures, presented on diverse object and sound classification tasks, demonstrate that the DAME approach provides improved explanation compared to other XAI methods.
Abstract : Test Time Adaptation (TTA) is a pivotal concept in machine learning, enabling models to perform well in real-world scenarios, where test data distribution differs from training. In this work, we propose a novel approach called pseudo Source guided Target Clustering (pSTarC) addressing the relatively unexplored area of TTA under real-world domain shifts. This method draws inspiration from target clustering techniques and exploits the source classifier for generating pseudo source samples. The test samples are strategically aligned with these pseudo source samples, facilitating their clustering and thereby enhancing TTA performance. pSTarC operates solely within the fully test-time adaptation protocol, removing the need for actual source data. Experimental validation on a variety of domain shift datasets, namely VisDA, Office-Home, DomainNet-126, CIFAR-100C verifies pSTarC’s effectiveness. This method exhibits significant improvements in prediction accuracy along with efficient computational requirements. Furthermore, we also demonstrate the universality of the pSTarC framework by showing its effectiveness for the continuous TTA framework.
Abstract : “A major challenge involved in automatic dysarthria severity classification for patients with Amyotrophic Lateral Sclerosis (ALS) is the difficulty to build a speech corpus which is large enough to train accurate and generalizable classifiers. To overcome this constraint, we employ transfer learning approaches, specifically, fine-tuning from an auxiliary task and multi-task learning. Input feature reconstruction and gender classification, on the same ALS speech dataset or other healthy speech corpora, are explored as the auxiliary tasks. We use temporal statistics of mel-frequency cepstral coefficients as the features and dense neural networks for performing the primary and auxiliary tasks. Experiments suggest that transfer learning aids severity classification with up to 11.03% absolute increase in the average classification accuracy as compared to direct single task
learning. The improvement is attributed mainly to better classification of the mild class than severe/normal classes.”
Abstract : Estimating translations (or locations) from directions (or unit vectors) has been of interest in many research areas like computer vision and sensor network localization. Due to dissimilarity in the input and output spaces, multiple challenges arise. In this poster, we will look into the sensitivity of the problem, i.e. change in output translations with small perturbations in the input directions. We study the conditions under which a set of directions is sensitive. We demonstrate the benefits of sensitivity analysis in structure-from-motion, an approach in computer vision to obtain 3D reconstructions of scenes given its images.
Abstract : With the rapid growth in uncrewed aircraft systems (UAS) being deployed, establishing a UAS Traffic Management (UTM) system for urban airspace is necessary. Existing UTMs designate preplanned deconflicted flight paths to UAS, enabling the UAS to deliver payloads. However, with stochastic delivery demand on these paths, UAS will likely experience congestion. We propose a congestion mitigation strategy that improves UAS safety in congested traffic streams. Following the strategy, UAS opts for alternative local paths surrounding the nominal path and avoids congested regions. The UAS reverts to the nominal path when congestion reduces. The strategy results in UAS traffic exploring and spreading to alternative adjacent routes when encountering congestion. We have developed queuing models to estimate the expected traffic spread for varying stochastic delivery demand at the source. We can accommodate any foreseen congestion by reserving the airspace around the nominal path beforehand. More traffic spread means more time delays. When multiple paths exist between a source and destination pair, we provide a path recommendation strategy for deploying UAS on a specific path. We model a resource allocation problem in a restless bandit framework with multiple users sharing paths as resources and traffic spread on these paths as metrics. The objective is to learn an optimal threshold policy based on the Whittle index that suggests that users should deploy UAS on the path with the highest Whittle index.
Abstract : We consider the problem of optimal unsignalized intersection management, wherein we seek to obtain safe and optimal trajectories, for a set of robots that arrive randomly and continually. This problem involves repeatedly solving a mixed integer program (with robot acceleration trajectories as decision variables) with different parameters, for which the computation time using a naive optimization algorithm scales exponentially with the number of robots and lanes. Hence, such an approach is not suitable for real-time implementation. In this work, we propose a solution framework that combines learning and sequential optimization. In particular, we propose an algorithm for learning a shared policy that given the traffic state information, determines the crossing order of the robots. Then, we optimize the trajectories of the robots sequentially according to that crossing order. This approach inherently guarantees safety at all times. We validate the performance of this approach using extensive simulations and compare our approach against various heuristics from the literature in various simulation settings. Our approach, on average, significantly outperforms the heuristics from the literature in various metrics like objective function, weighted average of crossing times and computation time. We also show through simulations that the computation time for our approach scales linearly with the number of robots (assuming all other factors are constant). We further implement the learnt policies on physical robots with a few modifications to the solution framework to address real-world challenges and establish its real-time implementability.