Risi Kondor (University of Chicago)

Title: ML foundations; Harmonic analysis & group representation theory

Abstract: Lie group equivariant networks have become the standard framework for modeling molecular systems and many systems of partial differential equations with neural networks, as well as a range of tasks in robotics and 3D vision. At the same time, equivariance to permutations has emerged as the crucial constraint in some other problems, specifically graph neural networks. In this talk I will present the general mathematical theory that encompasses equivariance to both the continuous and discrete symmetries, contrast the two settings, and specifically discuss the interaction between them in problems that possess both types of symmetry. I will talk about recent work on higher order permutation equivariance in graph neural networks, and specifically discuss computational aspects of implementing equivariant networks.