Abstract

A large class of problems in machine learning pertains to making sense of high dimensional and unlabelled data. The challenge lies in separating direct variable-variable interactions (e.g. cause and effect) and transitive correlations, as well as removing noise due to insufficient number of samples relative to the number of variables. In this talk, I will discuss an Ornstein-Zernike-like approach for data analysis that disentangles correlations in datasets using ideas from the theory of liquids. The Ornstein-Zernike closure is parameterised by deep learning, and a framework inspired by random matrix theory is used to remove finite sampling noise. I will illustrate this approach by applying it to problems such as ligand-based virtual screening and predicting protein function from sequence covariation.