Random Feature Stein Discrepancies

  • 2018-06-20 15:14:15
  • Jonathan H Huggins, Lester Mackey
  • 3

Abstract

Computable Stein discrepancies have been deployed for a variety ofapplications, including sampler selection in posterior inference, approximateBayesian inference, and goodness-of-fit testing. Existingconvergence-determining Stein discrepancies admit strong theoretical guaranteesbut suffer from a computational cost that grows quadratically in the samplesize. While linear-time Stein discrepancies have been proposed forgoodness-of-fit testing, they exhibit avoidable degradations in testingpower---even when power is explicitly optimized. To address these shortcomings,we introduce feature Stein discrepancies ($\Phi$SDs), a new family of qualitymeasures that can be cheaply approximated using importance sampling. We showhow to construct $\Phi$SDs that provably determine the convergence of a sampleto its target and develop high-accuracy approximations---random $\Phi$SDs(R$\Phi$SDs)---which are computable in near-linear time. In our experimentswith sampler selection for approximate posterior inference and goodness-of-fittesting, R$\Phi$SDs typically perform as well or better than quadratic-timeKSDs while being orders of magnitude faster to compute.

 

Quick Read (beta)

loading the full paper ...