Speaker Biography
Antonio Artés is a is a Professor at the Department of Signal Theory and Communications, Universidad Carlos III de Madrid, Spain. His research interests include signal processing and machine learning methods, and its application to human behavior characterization and health, with special focus on mental health.
Talk Abstract
Our digital footprint (behavioral data) is of high interest in healthcare but its heterogeneous nature makes it difficult to be employed in machine learning models. The main issue arises when we introduce factorized likelihood distributions over an arbitrary mix of continuous and discrete variables and posterior inference becomes non-trivial. In this talk, we present a solution based on Gaussian Process (GP) models. The idea is to link different non-linear mappings to the heterogeneous likelihood parameters, each one being sampled from a GP prior. To jointly model them, we assume a linear combination of an additional subset of latent functions. This is equivalent to place a multi-output GP prior over all the heterogeneous likelihood parameters. Additionally, inspired in conjugate-exponential methods, we expand the model to the case of streaming data, allowing the inference mechanism to propagate forward the uncertainty captured in the inducing points. The method is based on the posterior predictive GP equation and sparse approximations. It is robust up to thousands of iterations, e.g. updates of the model without re-observing past data samples, close to the performance level of filtering/signal processing methods.