CS Seminar: Kaiyi Ji

Description

Kaiyi Ji, an assistant professor in the Department of Computer Science and Engineering of the University at Buffalo, will give a talk titled "Efficient Bilevel Optimization for Machine Learning: Algorithms, Convergence, and Applications" for the Department of Computer Science. Ji is also an affiliated faculty member with the University at Buffalo's Institute for Artificial Intelligence and Data Science.

Abstract:

Bilevel optimization has emerged as a foundational framework in modern machine learning (ML)for developing principled computational tools across diverse domains, including meta-learning, automated ML, reinforcement learning, and robotics. This talk will explore recent advancements in bilevel optimization algorithms and their applications in machine learning. In the first part, Kaiyi Ji will introduce several efficient implicit gradient-based algorithms featuring flexible double- and single-loop structures. He will then present a novel adaptive tuning-free approach that significantly reduces hyperparameter tuning efforts while maintaining strong convergence guarantees. For the more challenging scenario involving non-unique lower-level solutions, Ji will discuss penalty-based methods that ensure provable convergence using only first-order gradient information. The second part of the talk will focus on two practical applications of bilevel optimization: coreset selection for machine learning and imperative learning for optimization and robotics. Finally, Ji will discuss promising directions for future research in theoretical and applied bilevel optimization and in other areas such as multi-objective learning and continual learning.