Efficient Bilevel Optimization and Application in Continual Learning

Kaiyi Ji
Assistant Professor
State University of New York at Buffalo
ECSE Topical Seminar
JEC 3117
Wed, December 06, 2023 at 4:00 PM

Bilevel optimization (BO) has received increasing attention in modern machine learning (ML), and has become a theoretical foundation for designing efficient computational tools for various ML areas such as meta-learning, autoML, fair ML, continual learning, and etc. In the first part of this talk, I will briefly introduce several recent BO applications in hyperparameter optimization and rehearsal based continual learning. In the second part, I will propose a novel stochastic bilevel optimization algorithm named stocBiO, which features a sample-efficient hypergradient estimation via Hessian-vector computations and automatic differentiation. I will then present the convergence analysis for stocBiO and discuss its application in coreset selection for rehearsal based continual learning by proposing a new bilevel problem formulation. I will finally discus the opportunities such as application of Hessian-free bilevel algorithms in large-scale continual learning, etc. 

Kaiyi Ji is currently an assistant professor at the Department of Computer Science and Engineering of the University at Buffalo (UB), and is also an affiliated faculty with the Institute for Artificial Intelligence and Data Science (IAD).  Dr. Ji was a postdoctoral research fellow at the Electrical Engineering and Computer Science Department of the University of Michigan, Ann Arbor, in 2022. Dr. Ji received the Ph.D. degree from the Electrical and Computer Engineering Department of The Ohio State University in December, 2021. Dr. Ji was a visiting student research collaborator at the department of Electrical Engineering, Princeton University. Previously Dr. Ji obtained the B.S. degree from the University of Science and Technology of China in 2016. Dr. Ji has worked at the intersection of optimization, machine learning and communication networks, on both the theory and application sides. His current major interest lies in bilevel optimization and its application to various machine learning areas including meta-learning, continual learning, hyperparameter optimization, etc. He received the prestigious Presidential Fellowship at OSU in 2020, and two NSF awards in 2023.