Dr. Anastasia Bizyaeva (AI Institute in Dynamic Systems, Univ. of Washington)

11/16/2023  3:10pm

Abstract: This work is motivated by the study of recurrent neural networks, networked systems, large-scale biological processes, and other systems that exhibit complex dynamic behavior. Mathematical models of these systems are often high in dimension, but exhibit low-dimensional effective behavior that is nontrivial, e.g. oscillatory or chaotic. We motivate, apply, and analyze an approach for model reduction of systems in Lur'e form that builds a low-dimensional model of the complex dynamics, leveraging data from simulations of the high-dimensional system. This method synthesizes insights from classic control-theoretic approaches with recent advances in data-driven methods. Effectiveness of this approach is illustrated by extracting a low-dimensional model from a trained chaotic recurrent neural network, and on a model of a biologically inspired network of oscillators.