CMStatistics 2022: Start Registration
View Submission - CMStatistics
B0874
Title: Finite element and graphical representations of Gaussian processes Authors:  Daniel Sanz-Alonso - University of Chicago (United States) [presenting]
Ruiyi Yang - Princeton University (United States)
Abstract: Gaussian processes (GPs) are popular models for random functions in computational and applied mathematics, statistics, machine learning and data science. However, GP methodology scales poorly to large data sets due to the need to factorize a dense covariance matrix. In spatial statistics, a standard approach to surmount this challenge is to represent Matern GPs using finite elements, obtaining an approximation with a sparse precision matrix. A new understanding of this approach will be given for regression and classification with large data sets, showing that under mild smoothness assumptions, the dimension of the matrices that need to be factorized can be reduced without hindering the estimation accuracy. The analysis balances finite element and statistical errors to show that there is a threshold beyond which further refining of the discretization increases the computational cost without improving the estimation accuracy. We will also introduce graphical representations of GPs to model random functions on high-dimensional point clouds, greatly expanding the important but limited scope of the finite element approach. We will show error bounds on the graphical representations, and study the associated posterior contraction in a semi-supervised learning problem.