Machine learning has driven many technological breakthroughs, yet training neural networks has become increasingly expensive. In this talk, I will present my research on developing theoretically-founded methods to improve the efficiency of machine learning, with a focus on optimization. Finding the best training algorithm and tuning its parameters is a costly but often necessary part of the training process. However, finding the best algorithm for particular optimization instances is a non-convex problem that is challenging for theory and practice. We will begin by describing meta-optimization, a framework for learning the best optimizer from problem instances. We will then discuss how meta-optimization can be formulated as a feedback control problem, and how recent advances in online control leads to provable methods for meta-optimization. We will conclude with how foundational tools in control theory can advance efficient architecture design, highlighting the expanding intersection between optimization, control theory, and machine learning.

Speaker bio: Xinyi Chen is a PhD candidate in the Department of Computer Science at Princeton University advised by Prof. Elad Hazan. She is also a research scientist at Google DeepMind. Her research is at the intersection of machine learning, optimization, and dynamical systems, in particular developing provably efficient methods for sequential decision-making and control. Previously, she obtained her undergraduate degree from Princeton in Mathematics, where she received the Middleton Miller Prize. Among other honors, she has received the NSF Graduate Research Fellowship and the Siebel Scholarship, and has been recognized by EECS Rising Stars at UC Berkeley and Rising Stars in Data Science at the University of Chicago.