Yes, but Did It Work?: Evaluating Variational Inference

  • 2018-02-07 17:25:36
  • Yuling Yao, Aki Vehtari, Daniel Simpson, Andrew Gelman
  • 54

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.

 

Introduction (beta)

None

 

Conclusion (beta)

None