Abstract
To make decisions based on a model fit by auto-encoding variational Bayes(AEVB), practitioners often let the variational distribution serve as asurrogate for the posterior distribution. This approach yields biased estimatesof the expected risk, and therefore poor decisions for two reasons. First, themodel fit by AEVB may yield biased statistics relative to the underlying datadistribution. Second, there may be strong discrepancies between the variationaldistribution and the posterior. We explore how fitting the variationaldistribution based on several objective functions other than the ELBO, whilecontinuing to fit the generative model based on the ELBO, affects the qualityof downstream decisions. For the probabilistic principal component analysismodel, we investigate how importance sampling error, as well as the biases inmodel parameter estimates, vary across several approximate posteriors when usedas proposal distributions. Our theoretical results suggest that a posteriorapproximation distinct from the variational distribution should be used formaking decisions. Motivated by these theoretical results, we propose learningseveral approximate proposals for the best model and combining them usingmultiple importance sampling for decision-making. In addition to toy examples,we present a full-fledged case study of single-cell RNA sequencing. In thischallenging instance of multiple hypothesis testing, our proposed approachsurpasses the current state of the art. The code and datasets are available athttp://github.com/PierreBoyeau/decision-making-vaes.