Search papers, labs, and topics across Lattice.
This paper introduces ReinMax-Rao and ReinMax-CV, two novel gradient estimators for discrete latent variables that build upon the ReinMax estimator. These estimators reduce variance by incorporating Rao-Blackwellisation and control variate techniques, respectively. Experiments on variational autoencoders with discrete latent spaces demonstrate that the proposed estimators outperform ReinMax.
ReinMax, a promising gradient estimator for discrete latent variables, gets a low-variance upgrade with Rao-Blackwellisation and control variates, boosting performance in discrete VAEs.
Machine learning models involving discrete latent variables require gradient estimators to facilitate backpropagation in a computationally efficient manner. The most recent addition to the Straight-Through family of estimators, ReinMax, can be viewed from a numerical ODE perspective as incorporating an approximation via Heun's method to reduce bias, but at the cost of high variance. In this work, we introduce the ReinMax-Rao and ReinMax-CV estimators which incorporate Rao-Blackwellisation and control variate techniques into ReinMax to reduce its variance. Our estimators demonstrate superior performance on training variational autoencoders with discrete latent spaces. Furthermore, we investigate the possibility of leveraging alternative numerical methods for constructing more accurate gradient approximations and present an alternative view of ReinMax from a simpler numerical integration perspective.
