Jiaxin Shi
Research Scientist, DeepMind, UK
Spectral Methods and Generative Modeling: A Unifying Perspective
Abstract
I will give an overview of our recent work that connects two major directions in unsupervised learning—spectral methods and generative models. Historically, these two classes of methods are often considered separately. In this talk, I will show that it is possible to connect them through a spectral representation of the data distribution gradients. This motivates several promising ideas for unsupervised learning, including score-based modelling and self-supervised learning with eigenfunctions. I will discuss the statistical principles behind these developments such as Stein’s method, and how we can leverage them to extend such methods to challenging domains with discrete structures and other constraints.
About Jiaxin Shi
Jiaxin Shi is a research scientist at Google DeepMind. Previously, he was a postdoctoral researcher at Stanford and Microsoft Research New England. He obtained B.E. and Ph.D. from Tsinghua University, advised by Jun Zhu. His research interests broadly involve probabilistic and algorithmic models for learning as well as the interface between them. Jiaxin served as an area chair for NeurIPS and AISTATS. He is a recipient of Microsoft Research PhD fellowship. His first-author paper was recognized by a NeurIPS 2022 outstanding paper award.