X.F. Gao
Langevin Dynamics (LD) have received considerable attention recently in the field of machine learning and computational statistics. LD has been proven to be powerful techniques for two closely-related tasks: 1) globally optimizing a non-convex objective function, and 2) sampling from a high-dimensional probability distribution. Langevin dynamics is based on the overdamped Langevin stochastic differential equation which is reversible in time. In this project, we aim to understand how breaking the reversibility could accelerate the Langevin dynamics for both optimization and sampling.