gconv

Exploiting Learned Symmetries in Group Equivariant Convolutions

We show that GConvs can be efficiently decomposed into depthwise separable convolutions while preserving equivariance properties and demonstrate improved performance and data efficiency on two datasets.