Abstract

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

 

Quick Read (beta)