Using Gradient Descent for Gaussian Process Prediction

• Activity Number: 104 - Advances in Bayesian Analysis of Computer Models
• Type: Topic Contributed
• Date/Time: Monday, August 8, 2022 : 8:30 AM to 10:20 AM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Matthew Plumlee*
• Companies: Northwestern University
• Keywords: kriging; computer experiments

Abstract

Gaussian processes are popular tools for Bayesian prediction that have been shown to be orders of magnitude more accurate that modern competitors on a host of prediction tasks. However, the computational cost of fitting them can be daunting. Inspired by the recent deployments of large scale optimization in deep learning, this talk illustrates how carefully written optimization problems can be used to replace the usual matrix decomposition used to fit Gaussian process predictors.