“Group Equivariant Convolutional Networks,” submitted to arXiv on February 24, 2016 by Taco Cohen and Max Welling, generalized the convolution operation so that networks could exploit symmetries beyond simple shifts. Ordinary convolutional networks are translation equivariant: shift the input and the feature maps shift correspondingly. Cohen and Welling extended this property to larger symmetry groups including rotations and reflections.
Their G-CNNs introduce a group convolution that shares weights not only across spatial positions but also across the transformations in a chosen discrete symmetry group, such as 90-degree rotations and mirror flips. This gives a substantially higher degree of weight sharing than a regular convolution layer, which means more expressive power for the same number of parameters and a model that automatically understands rotated or reflected versions of a pattern without having to learn them separately.
The practical payoff is reduced sample complexity: because the network builds in the symmetry rather than learning it from examples, it needs less data to reach a given accuracy. The authors demonstrated state-of-the-art results on standard image classification benchmarks at the time.
This paper is a cornerstone of equivariant deep learning. For domains where orientation should not change the answer, such as medical imaging, microscopy, satellite imagery, and molecular structures, building symmetry directly into the model is both more accurate and more data-efficient than hoping the network learns it.