Achieving equivariance in neural networks

There are several competing strategies for achieving equivariance in neural networks. In this talk, we are going to take a theoretical closer look at two of them: data augmentation and architecture restriction. Our main question is when the two strategies are equivalent. The analysis will reveal that a geometrical relation between the network architecture and the space of equivariant linear layers will imply a weak equivalence of the two strategies. This talk is based on joint work with Fredrik Ohlsson and Oskar Nordenfors.