Sampling with Langevin Algorithms in Continuous and Discrete Times

Title

Sampling with Langevin Algorithms in Continuous and Discrete Times

Speaker

Andre Wibisono - Yale University

Abstract

Sampling is a fundamental algorithmic task that appears in many applications. Many algorithms for sampling can be derived from the Langevin dynamics, which is a natural dynamics for sampling in continuous time. In this talk we will discuss two sampling algorithms, the Unadjusted Langevin Algorithms (ULA) and the Proximal Sampler, for sampling from target distribution under isoperimetry assumptions. We will survey recent results on the biased convergence guarantees of ULA. We will show how the Proximal Sampler can be viewed as a proximal discretization of the Langevin dynamics, and it gives unbiased convergence guarantees in discrete time that match the convergence guarantees of the Langevin dynamics in continuous time. This is joint work with Santosh Vempala, Yongxin Chen, Sinho Chewi, Adil Salim.

Bio

Andre Wibisono is an assistant professor in the Computer Science Department at Yale University. Andre did his postdoctoral work at Georgia Tech and at the University of Wisconsin-Madison. Andre received his PhD in EECS from UC Berkeley and his BS in Computer Science and in Mathematics from MIT. His research interests are in the design and analysis of algorithms for machine learning, in particular for optimization, sampling, and game dynamics.

When

June 26th, 15:00

Where

Room 322, DIBRIS, Via Dodecaneso 35