Presenter: Bin Hu, University of Illinois

Tuesday, October 18

8 AM Pacific / 9 AM Mountain / 10 AM Central / 11 AM Eastern for one hour

Abstract

Recent years have witnessed phenomenal accomplishments of deep learning in many application domains such as computer vision, natural language processing, Go, personalized healthcare, scientific discovery, and robotic manipulation. For real-world applications, deep learning models that lack robustness against adversarial attacks can run into unexpected catastrophic failures. To address such security/safety risks, Lipschitz network training has emerged as a principled approach for enhancing the robustness of deep learning models in the presence of adversarial attacks, and various techniques that can induce Lipschitz bounds on deep neural networks have been developed in a case-by-case manner.  This talk focuses on the connections between control theory and robust deep learning.  Specifically, we will tailor control-theoretic tools to develop a principled unified framework for enforcing Lipschtiz bounds and inducing robustness properties of deep neural networks in the presence of adversarial attacks. We will discuss how to leverage control-theoretic conditions in the form of semidefinite programs to develop various Lipschitz network structures which can be very useful for robust learning tasks. Several new insights on how to mitigate adversarial attacks in the deep network regime are also presented.

Biography

Bin Hu received the B.Sc. in Theoretical and Applied Mechanics from the University of Science and Technology of China in 2008, and received the M.S. in Computational Mechanics from Carnegie Mellon University in 2010. He received the Ph.D. in Aerospace Engineering and Mechanics at the University of Minnesota in 2016, advised by Peter Seiler. Between July 2016 and July 2018, he was a postdoctoral researcher in the Wisconsin Institute for Discovery at the University of Wisconsin-Madison. At Madison, he was working with Laurent Lessard and closely collaborating with Stephen Wright. He is currently an assistant professor in the Department of Electrical and Computer Engineering at the University of Illinois at Urbana-Champaign and affiliated with the Coordinated Science Laboratory. His research focuses on building fundamental connections between control and machine learning. In 2021, he received the NSF CAREER award and the Amazon research award.