NeurIPS 2020

Stein Self-Repulsive Dynamics: Benefits From Past Samples

Meta Review

This paper extends Langevin dynamics (LD) MCMC with an additive self-repulsion term which encourages rapid exploration of the target distribution. Strengths: - the idea is simple to implement and the paper is well-written - non-trivial and elegant theoretical analysis proves correctness Weaknesses: - empirically the method does not very strongly outperform Langevin dynamics - it is unclear how the method works in high dimensions