Seminars

View all Seminars  |  Download ICal for this event

The Sampling Problem Through The Lens of Optimization : Recent Advances and Insights

Series: Bangalore Theory Seminars

Speaker: Aniket Das Google Research

Date/Time: Nov 24 11:00:00

Location: CSA Seminar Hall (Room No. 254, First Floor)

Abstract:
The task of sampling from a probability measure on Rd whose density (w.r.t Lebesgue Measure) ?(x)?exp(??F(x)) is known only upto a normalizing constant, is a fundamental problem in High Dimensional Statistics, Theoretical Computer Science and Machine Learning. In this talk, I will discuss how sampling can be analyzed as an optimization-problem over the infinite-dimensional space of probability measures, equipped with the 2-Wasserstein metric from Optimal Transport. This perspective has spurred numerous breakthroughs in the field by allowing the transfer of familiar tools and techniques from the well-developed theory of continuous optimization. I will elucidate the efficacy of this paradigm by discussing the state-of-the-art analysis of two highly popular sampling algorithms: Stein Variational Gradient Descent (SVGD) and Stochastic Gradient Langevin Dynamics (SGLD).

This talk shall be based upon the following works:

Das and Nagaraj, Provably Fast Finite Particle Variants of SVGD via Virtual Particle Stochastic Approximation: Spotlight Paper at NeurIPS 2023; Oral Presentation at the Optimal Transport and Machine Learning Workshop 2023. [Link - Link

Das, Nagaraj and Raj, Utilising the CLT Structure in Stochastic Gradient-Based Sampling: Improved Analysis and Faster Algorithms : COLT 2023. [Link -Link


Microsoft Teams link:

Link

We are grateful to the Kirani family for generously supporting the theory seminar series

Hosts: Rameesh Paul, Rahul Madhavan, Rachana Gusain, KVN Sreenivas