Seminars
View all Seminars | Download ICal for this eventWhat is common to robots, proteins, genomics and video games?
Series: Department Seminar
Speaker: Kartic Subr University of Edinburgh
Date/Time: Feb 28 15:30:00
Location: CSA Seminar Hall (Room No. 254, First Floor)
Abstract:
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] Link
[2] Link
[3] Link
[4] Link
Speaker Bio:
Kartic Subr is a Royal Society University Research Fellow and Senior Lecturer (Assistant Prof) of Computer Graphics at the University of Edinburgh, UK. Kartics research addresses the development of approximate methods for computationally expensive procedures such as physical simulation. More recently he has also explored the use of neural approximations for their application to robotic manipulation. His previous employers include Heriot Watt University, Disney Research, University College London and INRIA-Grenoble. He obtained his PhD (2008, UC Irvine) under the supervision of James Arvo.
Host Faculty: Prof. Vijay Natarajan