Reproduce “A General Theory of Equivariant CNNs on Homogeneous Spaces” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Taco Cohen, Mario Geiger, Maurice Weiler

A general theory for Group Equivariant Convolutional Neural Networks (G-CNNs) operating on homogeneous spaces is presented, unifying existing architectures by modeling feature maps as sections of associated vector bundles. The work demonstrates that equivariant linear maps are convolutions with precisely characterized equivariant kernels, offering a systematic framework for classification and principled design of G-CNNs.

Key insights

  • Feature spaces in G-CNNs can be universally modeled as "fields," which are sections of associated vector bundles over homogeneous spaces, defined by a symmetry group G, a stabilizer H, and a representation .
  • Any G-equivariant linear map between these field types can be expressed as a convolution with a specific type of equivariant kernel, establishing convolution as a fundamental operation for symmetry-preserving networks.
  • The theory offers a systematic way to classify existing G-CNNs and design new ones by formalizing the components (G, H, ) that define their equivariance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.