Title: Building equivariant neural networks
Abstract: Equivariant neural networks, that is, neural networks which obey some prescribed symmetry, have become a central tool in applying techniques of deep learning in many scientific fields. However, the mathematical and computational underpinning of these networks may at first appear intimidating compared to the relative simplicity of classical architectures such as CNNs. In this talk, I will give an overview of some of the methods developed in the last couple of years to build neural networks which are equivariant to actions of compact Lie groups (with particular attention to SO3, the group of rotations in 3-dimension Euclidean space). I will discuss some of the trade-offs of the proposed methods, and the need for a better understanding of the types of data encountered in applications.
If you would like to attend, please email firstname.lastname@example.org for the Zoom link.