On Using Control Variates with Stochastic Approximation for Variational Bayes and its Connection to Stochastic Linear Regression

Abstract

Recently, we and several other authors have written about the possibilities of using stochastic approximation techniques for fitting variational approximations to intractable Bayesian posterior distributions. Naive implementations of stochastic approximation suffer from high variance in this setting. Several authors have therefore suggested using control variates to reduce this variance, while we have taken a different but analogous approach to reducing the variance which we call stochastic linear regression. In this note we take the former perspective and derive the ideal set of control variates for stochastic approximation variational Bayes under a certain set of assumptions. We then show that using these control variates is closely related to using the stochastic linear regression approximation technique we proposed earlier. A simple example shows that our method for constructing control variates leads to stochastic estimators with much lower variance compared to other approaches.

Cited by in corpus (8)

• Automatic Differentiation Variational Inference
• Automatic Variational Inference in Stan
• Generalized Variational Inference: Three arguments for deriving new Posteriors
• Determinantal Point Processes for Mini-Batch Diversification
• Fast and Scalable Spike and Slab Variable Selection in High-Dimensional Gaussian Processes
• CARMS: Categorical-Antithetic-REINFORCE Multi-Sample Gradient Estimator
• Double Control Variates for Gradient Estimation in Discrete Latent Variable Models
• A Divergence Bound for Hybrids of MCMC and Variational Inference and an Application to Langevin Dynamics and SGVI