Low rank matrix/tensor factorization using tools coming from Statistical physics

by Thibault Lesieur (Dip. di Fisica, La Sapienza, Roma)

Monday 29 Jan 2018, 14:30 → 16:00 Europe/Rome

Aula 5 (Dip. di Fisica - Edificio E. Fermi)

The fields of statistics is filled with methods that serve to  analyze and understand data,PCA method, kernel method, clustering method etc. Despite the successes of these methods there is still little guarantee on their performance. This is because there is no good model for what data coming from the real world looks like. In this talk I will analyze some of these problems through the prism of matrix and tensor factorization using Bayesian inference and tools coming from spin glass theory. The analysis of these systems will shed light on some inference problems that can be transformed in a matrix or tensor factorization problem.