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.
Description
Deep Unsupervised Learning using Nonequilibrium Thermodynamics
%0 Generic
%1 sohldickstein2015unsupervised
%A Sohl-Dickstein, Jascha
%A Weiss, Eric A.
%A Maheswaranathan, Niru
%A Ganguli, Surya
%D 2015
%K stat.ML
%T Deep Unsupervised Learning using Nonequilibrium Thermodynamics
%U http://arxiv.org/abs/1503.03585
%X 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.
@misc{sohldickstein2015unsupervised,
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.},
added-at = {2023-09-01T03:13:17.000+0200},
author = {Sohl-Dickstein, Jascha and Weiss, Eric A. and Maheswaranathan, Niru and Ganguli, Surya},
biburl = {https://www.bibsonomy.org/bibtex/29916befa78d8391c5c56f117e009f59b/aerover},
description = {Deep Unsupervised Learning using Nonequilibrium Thermodynamics},
interhash = {e4ec36cebbdfe482746365ef7fca7d96},
intrahash = {9916befa78d8391c5c56f117e009f59b},
keywords = {stat.ML},
note = {cite arxiv:1503.03585},
timestamp = {2023-09-01T03:13:17.000+0200},
title = {Deep Unsupervised Learning using Nonequilibrium Thermodynamics},
url = {http://arxiv.org/abs/1503.03585},
year = 2015
}