On Wednesday, April 10th, 2024, the seminar series hosted a talk by Dung Pham, PhD Researcher at the School of Computer Science and Statistics in Trinity College Dublin. Dung talked about approximate linear solvers for scalable statistical algorithms. Details for the talk are below.
Title
Approximate Linear Solvers for Scalable Statistical Algorithms: A Generalized Approach
Abstract
Solving large, sparse linear systems forms a critical foundation for numerous statistical methodologies and applications. The dimension of these systems frequently scales with the number of data points or the size of individual data vectors, reaching scales that render direct solutions computationally infeasible. This work introduces an approximate linear solver that employs the Krylov subspace method to address sparse,
high-dimensional linear systems. By exploiting inherent Kronecker structures arising during matrix construction, our proposed approach achieves generalization across a broad spectrum of applications.
In this seminar, we will discuss the motivations behind the solver’s development, its technical aspects, and its attempted application to various statistical problems. These problems include Cosmic Microwave Background source separation, efficient matrix- vector multiplication, and compression of auto-regressive language models.