Yes, but Did It Work?: Evaluating Variational Inference

TL;DR

This paper introduces two diagnostic algorithms, PSIS and VSBC, to evaluate the quality of variational inference approximations, addressing the challenge of identifying problems with these approximations.

Contribution

It presents novel diagnostic tools that improve the assessment of variational inference accuracy and reliability.

Findings

PSIS provides a goodness-of-fit measure for joint distributions.

VSBC evaluates the average performance of point estimates.

Both diagnostics help identify issues in variational approximations.

Abstract

While it's always possible to compute a variational approximation to a posterior distribution, it can be difficult to discover problems with this approximation. We propose two diagnostic algorithms to alleviate this problem. The Pareto-smoothed importance sampling (PSIS) diagnostic gives a goodness of fit measurement for joint distributions, while simultaneously improving the error in the estimate. The variational simulation-based calibration (VSBC) assesses the average performance of point estimates.