Reproduce “Generalized Shape Metrics on Neural Representations” 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

Alex H. Williams, Erin Kunz, Simon Kornblith, Scott W. Linderman

Researchers established generalized shape metrics for neural representations by applying statistical shape analysis and group theory, providing a rigorous framework for comparing networks while demonstrating that these metric spaces can be accurately embedded into Euclidean spaces to reveal interpretable structure in both biological and artificial networks.

Key insights

  • Representational dissimilarity can be rigorously defined as a proper metric by considering equivalence relations and group invariances over neural activations, providing a robust theoretical foundation for comparisons.
  • Despite their inherent non-Euclidean geometry, metric spaces of neural representations can be accurately approximated by Euclidean embeddings with low distortion, making them amenable to standard machine learning analyses.
  • The proposed generalized shape metrics reveal interpretable global structure in neural representations, reflecting meaningful biological organization in brain data and functional properties in deep neural networks.

Summary by alphaXiv.

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