Time：14:00-16:00, Friday, November 8 2024
Venue: E10-212
Speaker：Liming Wu, Harbin Institute of Technology
Biography: Wu Liming, professor at Universite Clermont-Auvergne since 1993 and at HIT since 2021, is a probabilist working on large deviations, functional inequalities and stochastic analysis. He has worked also at Wuhan University， and Academy of Mathematics and Systems Sciences in Chinese Academy of Sciences.
Title: Samplers of high dimensional Boltzmann distributions
Abstract: High dimensional Boltzmann distributions are basic models in statistical mechanics and in machine learning. However its computation is a NP-hard problem in the deterministic approaches: the curse of dimension.
In this talk I will present several current MCMC samplers such as Glauber dynamics, Gibbs samplers, and stochastic gradient descent. We will review some recent approaches for obtaining dimension-free estimates about the exponential convergence rate and the concentration inequalities, such as Poincare inequality, Gross' log-Sobolev inequality and Talagrand's transport inequality. Their applications in McKean-Vlasov equations and machine learning will be also presented.
