CHAI Research Talk: Matus Telgarsky
RIT's Center for Human Aware AI (CHAI) 2022 Spring Seminar Series
CHAI Research Talk: Matus Telgarsky, Ph.D., Assistant Professor, UIUC
Title: Approximation, Optimization, and Generalization in Deep Networks.
Abstract: This talk will survey three areas of deep learning theory, ranging from classical to modern results. The first question is: what functions can deep networks approximate? Classically, it was shown they can approximate anything, but this gives no hints about efficient approximations and architecture choices, so I will also present more recent results about benefits of depth and other architectural considerations. Secondly comes the question of generalization: why do deep networks achieve such good test error, despite achieving perfect fitting, or, arguably, overfitting? Here too I will survey classical results using generalization theory and point out their weaknesses, and point to more recent ideas, most of which are tied to optimization. The last topic is in fact optimization: since deep network training is non-convex, why do gradient methods work well? Here I will mainly discuss recent works, both discussing the near-initialization regime (the "neural tangent kernel"), and various preliminary works away from initialization, in the "feature learning regime".
Bio: Matus Telgarsky is an assistant professor at the University of Illinois, Urbana-Champaign. He studies machine learning theory, with a focus on deep learning theory and nonconvex optimization. He is an NSF CAREER award winner. He has significantly advanced the theory of deep learning, including the first proof demonstrating the benefits of depth in deep learning.
Prof. Telgarsky obtained his PhD in Computer Science from UCSD in 2013, under Sanjoy Dasgupta. While there, his research focused primarily upon optimization and statistical aspects of unconstrained and unregularized algorithms (e.g., boosting), and to a lesser extent, clustering. He then served as a postdoctoral researcher at Rutgers University and the University of Michigan, as well as a consulting researcher at Microsoft Research in New York City.
Event Snapshot
When and Where
Who
Open to the Public
| Cost | FREE |
Interpreter Requested?
Yes