Events

Approximate symmetries in machine learning

Speaker: Soledad Villar

Location: 60 Fifth Avenue, Room 650

Date: Friday, December 1, 2023

In this talk, we explain the different roles that symmetries and approximate symmetries can play in machine learning models. We define approximately equivariant graph neural networks and we show a bias-variance tradeoff when selecting the symmetries to enforce. We explain how to see equivariant functions as gradients of invariant functions, and we show how to use these ideas in self-supervised learning.