Title: Median of means estimation for high dimensional time series
Authors: Stephane Chretien - NPL (United Kingdom) [presenting]
Abstract: Linear dynamical systems with generalized linear observation models are considered. These models are traditionally estimated using maximum likelihood or moment based methods. Both approaches need in particular to tackle the joint estimation of the latent states which model the dependencies, and the transition parameters which model the dynamics. Recent work proposes to address the estimation problem using optimised matrix factorisation. Our contribution extends the matrix factorisation approach to the setting where outliers may be present, using Median of Means (MoM)-based robustification.