Discrete-time diffusion-based generative models and score matching methods
have shown promising results in modeling high-dimensional image data. Recently,
Song et al. (2021) show that diffusion processes that transform data into noise
can be reversed via learning the score function, i.e. the gradient of the
log-density of the perturbed data. They propose to plug the learned score
function into an inverse formula to define a generative diffusion process.
Despite the empirical success, a theoretical underpinning of this procedure is
still lacking. In this work, we approach the (continuous-time) generative
diffusion directly and derive a variational framework for likelihood
estimation, which includes continuous-time normalizing flows as a special case,
and can be seen as an infinitely deep variational autoencoder. Under this
framework, we show that minimizing the score-matching loss is equivalent to
maximizing a lower bound of the likelihood of the plug-in reverse SDE proposed
by Song et al. (2021), bridging the theoretical gap.
Description
[2106.02808] A Variational Perspective on Diffusion-Based Generative Models and Score Matching
%0 Journal Article
%1 huang2021variational
%A Huang, Chin-Wei
%A Lim, Jae Hyun
%A Courville, Aaron
%D 2021
%K bayesian optimization variational
%T A Variational Perspective on Diffusion-Based Generative Models and Score
Matching
%U http://arxiv.org/abs/2106.02808
%X Discrete-time diffusion-based generative models and score matching methods
have shown promising results in modeling high-dimensional image data. Recently,
Song et al. (2021) show that diffusion processes that transform data into noise
can be reversed via learning the score function, i.e. the gradient of the
log-density of the perturbed data. They propose to plug the learned score
function into an inverse formula to define a generative diffusion process.
Despite the empirical success, a theoretical underpinning of this procedure is
still lacking. In this work, we approach the (continuous-time) generative
diffusion directly and derive a variational framework for likelihood
estimation, which includes continuous-time normalizing flows as a special case,
and can be seen as an infinitely deep variational autoencoder. Under this
framework, we show that minimizing the score-matching loss is equivalent to
maximizing a lower bound of the likelihood of the plug-in reverse SDE proposed
by Song et al. (2021), bridging the theoretical gap.
@article{huang2021variational,
abstract = {Discrete-time diffusion-based generative models and score matching methods
have shown promising results in modeling high-dimensional image data. Recently,
Song et al. (2021) show that diffusion processes that transform data into noise
can be reversed via learning the score function, i.e. the gradient of the
log-density of the perturbed data. They propose to plug the learned score
function into an inverse formula to define a generative diffusion process.
Despite the empirical success, a theoretical underpinning of this procedure is
still lacking. In this work, we approach the (continuous-time) generative
diffusion directly and derive a variational framework for likelihood
estimation, which includes continuous-time normalizing flows as a special case,
and can be seen as an infinitely deep variational autoencoder. Under this
framework, we show that minimizing the score-matching loss is equivalent to
maximizing a lower bound of the likelihood of the plug-in reverse SDE proposed
by Song et al. (2021), bridging the theoretical gap.},
added-at = {2021-06-08T15:30:39.000+0200},
author = {Huang, Chin-Wei and Lim, Jae Hyun and Courville, Aaron},
biburl = {https://www.bibsonomy.org/bibtex/29b29bd97fef0a534e874e730932f078b/kirk86},
description = {[2106.02808] A Variational Perspective on Diffusion-Based Generative Models and Score Matching},
interhash = {504d93e60c994c1d598bb59114866ba4},
intrahash = {9b29bd97fef0a534e874e730932f078b},
keywords = {bayesian optimization variational},
note = {cite arxiv:2106.02808},
timestamp = {2021-06-08T15:30:39.000+0200},
title = {A Variational Perspective on Diffusion-Based Generative Models and Score
Matching},
url = {http://arxiv.org/abs/2106.02808},
year = 2021
}