Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models
(Submitted on 21 Mar 2017 (v1), last revised 6 Nov 2017 (this version, v4))
Abstract: Learning in models with discrete latent variables is challenging due to high variance gradient estimators. Generally, approaches have relied on control variates to reduce the variance of the REINFORCE estimator. Recent work (Jang et al. 2016, Maddison et al. 2016) has taken a different approach, introducing a continuous relaxation of discrete variables to produce low-variance, but biased, gradient estimates. In this work, we combine the two approaches through a novel control variate that produces low-variance, \emph{unbiased} gradient estimates. Then, we introduce a modification to the continuous relaxation and show that the tightness of the relaxation can be adapted online, removing it as a hyperparameter. We show state-of-the-art variance reduction on several benchmark generative modeling tasks, generally leading to faster convergence to a better final log-likelihood.
Submission history
From: George Tucker [view email][v1] Tue, 21 Mar 2017 18:05:31 GMT (598kb,D)
[v2] Sat, 22 Apr 2017 11:04:12 GMT (1238kb,D)
[v3] Thu, 8 Jun 2017 20:54:49 GMT (6080kb,D)
[v4] Mon, 6 Nov 2017 17:50:34 GMT (6099kb,D)
Link back to: arXiv, form interface, contact.