43 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Functional Stochastic Localization
UT Austin
Recording
Abstract
Stochastic localization has become a tool for high-dimensional geometry, sampling and diffusion-model analysis. Motivated by mirror descent and non-Euclidean sampling, this talk generalizes the process by replacing Gaussian regularization with multiples of log-Laplace transforms. It introduces the construction, its analysis, applications and remaining questions.