Learning rotation-agnostic representations via Group Equivariant VAEs

An emerging field in representation learning involves the study of group-equivariant neural networks, that leverage concepts from group representation theory to design neural architectures that can exploit discrete and continuous symmetries to produce more general representations. Following this direction, in this work we demonstrate how an image embedding agnostic to rotations can be naturally obtained by training a variational autoencoder (S-GVAE) equipped with a Group equivariant Convolutional Neural Network (G-CNN) encoder.