Friday, February 28, 2020 | 3:15PM | Yost 306

Title: Graph Representations of Gaussian Fields

Abstract

In this talk I will derive a graph representation of non-stationary Matérn fields that stems from the SPDE approach. Graph representations are natural alternatives to finite element and finite difference representations used in spatial statistics and are useful prior models in several inverse problems and machine learning applications. I will establish continuum limit guarantees for these prior models and highlight that they allow to naturally define a Markov random field approximation leading to a reduced computational cost by exploiting sparsity. I will also show continuum limit results for various Bayesian inverse problems and discuss the design of MCMC algorithms for graph-based inverse problems.

*Light refreshments will be served prior to the talk.