Reproduce “Latent Functional Maps: a spectral framework for representation alignment” 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 Fumero, Marco Pegoraro, Valentino Maiorca, Francesco Locatello, Emanuele Rodolà

Researchers introduce Latent Functional Maps (LFM), a spectral framework for aligning neural network representations by establishing correspondences between function spaces defined on latent manifolds. This method improved performance in zero-shot model stitching and cross-modal retrieval, demonstrating efficiency with limited supervision.

Key insights

  • Modeling the relationship between *function spaces* defined on latent manifolds offers a more flexible and interpretable approach to representation alignment than direct sample-to-sample mapping.
  • A computed functional map naturally serves as an interpretable similarity measure between latent spaces, quantifying deviations from isometry.
  • The framework enables robust transfer of representations and discovery of correspondences between spaces with minimal to no explicit supervision, bridging unsupervised, weakly supervised, and supervised settings.

Summary by alphaXiv.

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