NeurIPS 2020

VarGrad: A Low-Variance Gradient Estimator for Variational Inference

Meta Review

This paper presents a novel analysis of an existing algorithm, showing that it can be reinterpreted. In particular, take the leave-one-out control variate to reduce the variance of the score-function estimator. This can be seen as a "naive" gradient of a new log-variance loss. This appears to be a novel and valuable contribution.