Generative Adversarial Learning of Markov Chains

Jiaming Song, Shengjia Zhao, Stefano Ermon

TL;DR: We can train Markov Chains with an adversarial network.

Abstract: We investigate generative adversarial training methods to learn a transition operator for a Markov chain, where the goal is to match its stationary distribution to a target data distribution. We propose a novel training procedure that avoids sampling directly from the stationary distribution, while still capable of reaching the target distribution asymptotically. The model can start from random noise, is likelihood free, and is able to generate multiple distinct samples during a single run. Preliminary experiment results show the chain can generate high quality samples when it approaches its stationary, even with smaller architectures traditionally considered for Generative Adversarial Nets.

Keywords: Deep learning, Unsupervised Learning

Conflicts: stanford.edu, tsinghua.edu.cn, duke.edu