Current browse context:
stat.ML
Change to browse by:
References & Citations
Statistics > Machine Learning
Title: VarGrad: A Low-Variance Gradient Estimator for Variational Inference
(Submitted on 20 Oct 2020 (v1), last revised 29 Oct 2020 (this version, v2))
Abstract: We analyse the properties of an unbiased gradient estimator of the ELBO for variational inference, based on the score function method with leave-one-out control variates. We show that this gradient estimator can be obtained using a new loss, defined as the variance of the log-ratio between the exact posterior and the variational approximation, which we call the $\textit{log-variance loss}$. Under certain conditions, the gradient of the log-variance loss equals the gradient of the (negative) ELBO. We show theoretically that this gradient estimator, which we call $\textit{VarGrad}$ due to its connection to the log-variance loss, exhibits lower variance than the score function method in certain settings, and that the leave-one-out control variate coefficients are close to the optimal ones. We empirically demonstrate that VarGrad offers a favourable variance versus computation trade-off compared to other state-of-the-art estimators on a discrete VAE.
Submission history
From: Lorenz Richter [view email][v1] Tue, 20 Oct 2020 16:46:01 GMT (524kb,D)
[v2] Thu, 29 Oct 2020 10:27:27 GMT (525kb,D)
Link back to: arXiv, form interface, contact.