Statistics and Computational Methods Seminar Series - Spring 2025

Speaker: Reinhard Furrer (Universität Zürich)

Title: Spatial Statistics Meets Machine Learning: On Approximations and "Exact" Methods for Gaussian Processes

Abstract

User-friendly and scalable methods for Gaussian processes are in high demand across both classical spatial statistics and machine learning. Despite parallel progress over the past decade, cross-fertilization between the two communities has remained surprisingly limited — each field developing its own vocabulary, methods, and benchmarks largely in isolation. This talk surveys recent developments that bridge these two perspectives, with a focus on two contributions. The first is an iterative inference framework for full-scale GP approximations, paired with the fully independent training conditional (FITC) as a preconditioner. This combination substantially accelerates conjugate gradient solvers and stabilizes stochastic likelihood and variance estimation. The second is the Vecchia-inducing-point full-scale (VIF) approximation, a hybrid method that unifies global inducing points with local Vecchia conditioning. A recurring theme throughout is the tension between computational tractability and statistical rigor — and how viewing familiar ML methods through a statistical lens (or vice versa) can yield both new insights and practical gains.
 

Link streaming: https://teams.microsoft.com/meet/3719029318610?p=BHk4zdfy61myswxMbu