Constant-Time Predictive Distributions for Gaussian Processes

  • 2018-06-20 16:39:16
  • Geoff Pleiss, Jacob R. Gardner, Kilian Q. Weinberger, Andrew Gordon Wilson
  • 1

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 sped up GP marginal likelihood and posteriormean computations, leaving posterior covariance estimation and sampling as theremaining computational bottlenecks. In this paper we address theseshortcomings by using the Lanczos algorithm to rapidly approximate thepredictive covariance matrix. Our approach, which we refer to as LOVE (LanczOsVariance Estimates), substantially improves time and space complexity. In ourexperiments, LOVE computes covariances up to 2,000 times faster and drawssamples 18,000 times faster than existing methods, all without sacrificingaccuracy.

 

Quick Read (beta)

loading the full paper ...