New Particle-Based Variational Inferences Methods by a Discrete Energetic Variational Approach

Speaker: 

Jiuhai Chen, Ph.D. candidate (AMAT) Illinois Tech

Description: 

In this work, we present a new particle-based VI methods for Bayesian inference, inspired by a discrete energetic variational approach. Our starting point is a regularized energy dissipation law, then a discrete energetic variational approach gives as an interacting particle system (semi-discrete equation) that preserves the variational structure at a semi-discrete level. By solving the semi-discrete equation implicitly, our algorithms decrease the KL-divergence in each iteration and push the probability density to the target distribution efficiently. Empirical studies are performed on different examples, including real world application, indicating our method is competitive with existing methods. This is a joint work with Yiwei Wang, Lulu Kang and Chun Liu.

Event Topic:

Computational Mathematics & Statistics