GPU-accelerated block solvers for spatial-temporal Bayesian inference

You can join here

Speaker: 
Lisa Gaedke-Merzhäuser

Abstract:
Greater data availability and higher-dimensional model parameter spaces, that are especially prevalent when dealing with spatial-temporal data, lead to a growing demand in performing larger-scale Bayesian inference tasks. In this seminar talk we want to explore how methods from the field of high-performance computing can be employed to meet this demand. We will particularly focus on the methodology of integrated nested Laplace approximations (INLA), a popular framework for performing approximate Bayesian inference, that is applicable to a large class of Bayesian additive models. We will discuss how GPU-accelerated block solvers, that exploit the sparsity pattern of the arising matrices, can be employed to carry out INLA’s computational kernel operations. Additionally, we explore different opportunities for shared and distributed memory parallelism taking modern GPGPU node architectures into account.

Biography:
Lisa is a third-year PhD student at USI working under the supervision of Prof. Olaf Schenk. Her research focuses on combining statistical learning techniques with methods from high-performance computing. She holds a MSc degree in Mathematics from Freie Universität Berlin, Germany.