Metropolis-Hastings Stein Variational Gradient Descent (MET-SVGD)

This website contains materials introducing Metropolis-Hastings SVGD (MET-SVGD), a novel variational inference framework designed for efficient and scalable entropy estimation. MET-SVGD builds on ideas from Stein Variational Gradient Descent (SVGD), parametric variational inference (P-VI), and Metropolis-Hastings (MH), combining their strengths to offer a flexible and efficient approach to approximating complex distributions.

Besides an introduction to SVGD, variational inference, and related topics, we also provide documentation and code examples for the library accompanying the project.

• Section 1: Introduction

• Section 2: SVGD

• Section 3: The SVGD-Induced Density

• Section 4: MET-SVGD Optimizations

• Section 5: Results

• Section 6: Library Documentation

Reference¶

1. Messaoud, S., Mokeddem, B., Xue, Z., Pang, L., An, B., Chen, H., & Chawla, S. S²AC: Energy-based reinforcement learning with Stein soft actor critic. ICLR 2024.

2. Messaoud, S., Charni, S., Bouazza, E., Pourghasemi Fatideh, A., & Bensmail, H. Particles don’t care about Z: Towards scaling entropy estimation of unnormalized densities. ICML 2026.

Code¶

The accompanying library is available on GitHub: SafaMessaoud/MET-SVGD-Variational-Inference-With-SVGD.

Website prepared by Skander Charni.