Maciej Koch-Janusz, Haiqu Inc. & University of Zurich

Leuchs-Russell-Auditorium, A.1.500, Staudtstr. 2

Location details

Abstract:

In complex statistical systems we often seek to identify the relevant degrees of freedom out of high-dimensional raw configurational data, or more formally, the RG-relevant operators. In dynamical systems we seek model reductions in terms of a handful of variables most predictive of the future dynamics. I will discuss how both can be re-cast as the problem of finding the optimal compression preserving information about a suitably defined relevance variable, and how the relevant degrees of freedom are the solutions to a corresponding Information Bottleneck (IB) problem. This gives rise to an efficient numerical algorithm using contrastive learning. I will demonstrate applications to coarse-graining and extracting the operator content of classical and quantum statistical models and dynamical systems.