1 of 1

A-NICE-MC: Adversarial Training for MCMC

Source: https://github.com/ermongroup/a-nice-mc

How to design a Markov chain kernel that converges quickly to a given stationary distribution?

We propose a likelihood free method that

• Learns efficient domain specific Markov chain kernels parameterized by neural networks;
• Satisfies detailed balance by exploiting the volume preserving property of NICE[1];
• Allows fully-differentiable training though MCMC kernel is not.

Jiaming Song, Shengjia Zhao, Stefano Ermon

Sampling a distribution with 5 distinct rings.

[1] NICE: Non-linear Independent Components Estimation, Dinh et al.

Ermon Group, Stanford Artificial Intelligence Laboratory

Sampling faces through a (slow) Markov chain.