Discussion Lead: Philip Greengard [Columbia]
Topic: Equispaced Fourier representations for efficient Gaussian process regression
Abstract: Over the last couple of decades, a large number of numerical methods have been introduced for efficiently performing Gaussian process regression. Most of these methods rely on fast inversion of a covariance matrix whose size scales with the number of data points of the problem. Generally, these methods start to break down for problems with over a million data points, usually due to memory constraints, accuracy issues, or long compute times. In this talk, we describe a novel approach to Gaussian process regression that relies on efficient Fourier representations of Gaussian processes. Combining these representations with fast algorithms, we perform accurate and efficient Gaussian process regression for massive-scale problems in 1, 2, and 3 dimensions.