Reproduce “On Universality of Deep Equivariant Networks” 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

Marco Pacini, Mircea Petrache, Bruno Lepri, Shubhendu Trivedi, Robin Walters

Pacini et al. establish a comprehensive theoretical framework for the universality of deep invariant and equivariant networks, demonstrating how network depth and specific readout layers contribute to approximating symmetry-compatible functions. Their work introduces “entry-wise separability” as a refined criterion for accurately characterizing the expressive power of equivariant models.

Key insights

  • Standard separability is inadequate for characterizing the expressive power of equivariant networks; the novel "entry-wise separability" is crucial for a precise understanding.
  • The addition of a fully connected readout layer is a decisive mechanism for invariant networks to achieve universality, resolving prior issues where separation power did not imply approximation power.
  • Deep equivariant networks attain entry-wise separation-constrained universality either when their entry-wise separation relations stabilize with sufficient depth or when equipped with specific equivariant readout layers (e.g., width-1 convolutional filters).

Summary by alphaXiv.

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