Constant-Time Predictive Distributions for Gaussian Processes

  • 2018-03-19 17:49:21
  • Geoff Pleiss, Jacob R. Gardner, Kilian Q. Weinberger, Andrew Gordon Wilson
  • 0

Abstract

One of the most compelling features of Gaussian process (GP) regression isits ability to provide well calibrated posterior distributions. Recent advancesin inducing point methods have drastically sped up marginal likelihood andposterior mean computations, leaving posterior covariance estimation andsampling as the remaining computational bottlenecks. In this paper we addressthis shortcoming by using the Lanczos decomposition algorithm to rapidlyapproximate the predictive covariance matrix. Our approach, which we refer toas LOVE (LanczOs Variance Estimates), substantially reduces the time and spacecomplexity over any previous method. In practice, it can compute predictivecovariances up to 2,000 times faster and draw samples 18,000 time faster thanexisting methods, all without sacrificing accuracy.

 

Quick Read (beta)

loading the full paper ...