Reproduce “Topological Blindspots: Understanding and Extending Topological Deep Learning Through the Lens of Expressivity” 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

Yam Eitan, Yoav Gelberg, Guy Bar-Shalom, Fabrizio Frasca, Michael Bronstein, Haggai Maron

A study on Topological Deep Learning establishes the expressivity limitations of Higher-Order Message Passing (HOMP) models, showing they fail to capture fundamental topological and metric invariants of complex structures. It introduces Scalable Multi-Cellular Networks (SMCN), an architecture designed to mitigate these "topological blindspots," achieving superior performance on new topological benchmarks and various graph prediction tasks.

Key insights

  • Higher-Order Message Passing (HOMP) models exhibit "topological blindspots," being unable to distinguish combinatorial complexes based on fundamental properties like diameter, orientability, planarity, and homology groups.
  • Multi-Cellular Networks (MCN), by leveraging multi-cellular cochain spaces and equivariant linear layers, achieve full expressivity, capable of distinguishing any pair of non-isomorphic combinatorial complexes.
  • The Scalable Multi-Cellular Network (SMCN) architecture effectively mitigates many of HOMP's topological expressivity limitations and can leverage higher-order information while maintaining practical computational efficiency.

Summary by alphaXiv.

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