Jingfeng Wu (UC Berkeley): Statistics and DSI Joint Colloquium

Please join us for a Statistics and DSI joint colloquium.

Thursday, February 5
4:00pm –5:00pm
Data Science Institute, Room 105
5460 S University Ave, Chicago, IL 60615

Title: Towards a Less Conservative Theory of Machine Learning: Unstable Optimization and Implicit Regularization

Abstract: Deep learning’s empirical success challenges the “conservative” nature of classical optimization and statistical learning theories. Classical theory mandates small stepsizes for training stability and explicit regularization for complexity control. Yet, deep learning leverages mechanisms that thrive beyond these traditional boundaries. In this talk, I present a research program dedicated to building a less conservative theoretical foundation by demystifying two such mechanisms:

1. Unstable Optimization: I show that large stepsizes, despite causing local oscillations, accelerate the global convergence of gradient descent (GD) in overparameterized logistic regression.

2. Implicit Regularization: I show that the implicit regularization of early-stopped GD statistically dominates explicit $\ell_2$-regularization across all linear regression problem instances.

I further showcase how the theoretical principles lead to practice-relevant algorithmic designs (such as Seesaw for reducing serial steps in large language model pretraining). I conclude by outlining a path towards a rigorous understanding of modern learning paradigms.

Bio: I am a postdoc fellow at the Simons Institute at UC Berkeley hosted by Peter Bartlett and Bin Yu. I am a part of the NSF/Simons Collaboration on the Theoretical Foundations of Deep Learning. I received my Ph.D. in Computer Science at Johns Hopkins University, advised by Vladimir Braverman.

I work on the theory and algorithms for machine learning. I am interested in deep learning theory, optimization, and statistical learning.