Exploiting Learned Symmetries in Group Equivariant Convolutions
This work addresses computational bottlenecks in equivariant neural networks for researchers and practitioners, though it appears incremental as it builds on existing GConv frameworks.
The authors tackled the parameter and computational inefficiency of Group Equivariant Convolutions (GConvs) by analyzing learned filter symmetries, showing they can be decomposed into depthwise separable convolutions while preserving equivariance, resulting in improved performance and data efficiency on two datasets.
Group Equivariant Convolutions (GConvs) enable convolutional neural networks to be equivariant to various transformation groups, but at an additional parameter and compute cost. We investigate the filter parameters learned by GConvs and find certain conditions under which they become highly redundant. 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. All code is publicly available at github.com/Attila94/SepGrouPy.