Abstract

We propose a novel particle-based variational inference method designed towork with multimodal distributions. Our approach, referred to as Branched SteinVariational Gradient Descent (BSVGD), extends the classical Stein VariationalGradient Descent (SVGD) algorithm by incorporating a random branching mechanismthat encourages the exploration of the state space. In this work, a theoreticalguarantee for the convergence in distribution is presented, as well asnumerical experiments to validate the suitability of our algorithm. Performancecomparisons between the BSVGD and the SVGD are presented using the Wassersteindistance between samples and the corresponding computational times.

 

Quick Read (beta)