Speaker: Dr. Andre Wibisono, Yale
Focus: Machine Learning, Optimization

Title: “Mixing Times of the Langevin Dynamics and the Proximal Sampler for Sampling under Isoperimetry”

Abstract: Sampling is a fundamental algorithmic task with intricate connections with optimization, for example via the perspective of sampling as optimization in the space of distributions. In this talk, we discuss the Langevin dynamics, which is the natural greedy dynamics for sampling in continuous time, and its discrete-time implementations including the Proximal Sampler, a new sampling algorithm which is unbiased. We survey their mixing time guarantees under isoperimetry including the log-Sobolev inequality, and highlight their similarities with the convergence guarantees from optimization. We review the analysis techniques which proceed via establishing strong data processing inequalities for the algorithms when viewed as information-theoretic channels.

The KSU Mathematics Colloquium is a seminar series that brings in researchers from other universities to present results on all areas of Mathematics. The talks are targeted at general audiences of students and professors.

Dial-In Information
Stream Link
Passcode: 5VUHje