Deep Unsupervised Learning using Nonequilibrium Thermodynamics

TL;DR

This paper introduces a novel deep generative modeling approach inspired by nonequilibrium thermodynamics, which combines flexibility and tractability through a diffusion process that gradually destroys and then restores data structure.

Contribution

The authors develop a diffusion-based generative modeling framework that achieves both high flexibility and computational tractability, enabling efficient learning, sampling, and inference.

Findings

Allows rapid learning and sampling from deep models with thousands of layers

Enables computation of conditional and posterior probabilities efficiently

Provides an open source implementation of the method

Abstract

A central problem in machine learning involves modeling complex data-sets using highly flexible families of probability distributions in which learning, sampling, inference, and evaluation are still analytically or computationally tractable. Here, we develop an approach that simultaneously achieves both flexibility and tractability. The essential idea, inspired by non-equilibrium statistical physics, is to systematically and slowly destroy structure in a data distribution through an iterative forward diffusion process. We then learn a reverse diffusion process that restores structure in data, yielding a highly flexible and tractable generative model of the data. This approach allows us to rapidly learn, sample from, and evaluate probabilities in deep generative models with thousands of layers or time steps, as well as to compute conditional and posterior probabilities under the learned model. We additionally release an open source reference implementation of the algorithm.