Despite the successes in capturing continuous distributions, the application
of generative adversarial networks (GANs) to discrete settings, like natural
language tasks, is rather restricted. The fundamental reason is the difficulty
of back-propagation through discrete random variables combined with the
inherent instability of the GAN training objective. To address these problems,
we propose Maximum-Likelihood Augmented Discrete Generative Adversarial
Networks. Instead of directly optimizing the GAN objective, we derive a novel
and low-variance objective using the discriminator's output that follows
corresponds to the log-likelihood. Compared with the original, the new
objective is proved to be consistent in theory and beneficial in practice. The
experimental results on various discrete datasets demonstrate the effectiveness
of the proposed approach.
%0 Generic
%1 che2017maximumlikelihood
%A Che, Tong
%A Li, Yanran
%A Zhang, Ruixiang
%A Hjelm, R Devon
%A Li, Wenjie
%A Song, Yangqiu
%A Bengio, Yoshua
%D 2017
%K GAN discrete
%T Maximum-Likelihood Augmented Discrete Generative Adversarial Networks
%U http://arxiv.org/abs/1702.07983
%X Despite the successes in capturing continuous distributions, the application
of generative adversarial networks (GANs) to discrete settings, like natural
language tasks, is rather restricted. The fundamental reason is the difficulty
of back-propagation through discrete random variables combined with the
inherent instability of the GAN training objective. To address these problems,
we propose Maximum-Likelihood Augmented Discrete Generative Adversarial
Networks. Instead of directly optimizing the GAN objective, we derive a novel
and low-variance objective using the discriminator's output that follows
corresponds to the log-likelihood. Compared with the original, the new
objective is proved to be consistent in theory and beneficial in practice. The
experimental results on various discrete datasets demonstrate the effectiveness
of the proposed approach.
@misc{che2017maximumlikelihood,
abstract = {Despite the successes in capturing continuous distributions, the application
of generative adversarial networks (GANs) to discrete settings, like natural
language tasks, is rather restricted. The fundamental reason is the difficulty
of back-propagation through discrete random variables combined with the
inherent instability of the GAN training objective. To address these problems,
we propose Maximum-Likelihood Augmented Discrete Generative Adversarial
Networks. Instead of directly optimizing the GAN objective, we derive a novel
and low-variance objective using the discriminator's output that follows
corresponds to the log-likelihood. Compared with the original, the new
objective is proved to be consistent in theory and beneficial in practice. The
experimental results on various discrete datasets demonstrate the effectiveness
of the proposed approach.},
added-at = {2017-08-18T10:29:17.000+0200},
author = {Che, Tong and Li, Yanran and Zhang, Ruixiang and Hjelm, R Devon and Li, Wenjie and Song, Yangqiu and Bengio, Yoshua},
biburl = {https://www.bibsonomy.org/bibtex/2689951de5495959e2d43ef2eb1050199/daschloer},
description = {Maximum-Likelihood Augmented Discrete Generative Adversarial Networks},
interhash = {a9a8f5f6bcf9fe5f92a7cd6c454ae4db},
intrahash = {689951de5495959e2d43ef2eb1050199},
keywords = {GAN discrete},
note = {cite arxiv:1702.07983Comment: 11 pages, 3 figures},
timestamp = {2017-10-04T16:28:59.000+0200},
title = {Maximum-Likelihood Augmented Discrete Generative Adversarial Networks},
url = {http://arxiv.org/abs/1702.07983},
year = 2017
}