Reproduce “Implicit Gaussian process representation of vector fields over arbitrary latent manifolds” 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

Robert L. Peach, Matteo Vinao-Carl, Nir Grossman, Michael David, Emma Mallas, David Sharp, Paresh A. Malhotra, Pierre Vandergheynst and 1 more

This research from University Hospital Würzburg, Imperial College London, and EPFL introduces Riemannian Manifold Vector-field Gaussian Process (RVGP), a framework extending Gaussian processes to model vector fields over unknown latent Riemannian manifolds. RVGP utilizes graph-based approximations and the connection Laplacian for intrinsic geometric encoding, enabling accurate super-resolution and inpainting of vector fields, and significantly improving classification accuracy for Alzheimer's disease from low-density EEG recordings.

Key insights

  • The connection Laplacian's spectral components serve as an effective intrinsic positional encoding, enabling Gaussian Processes to model vector fields even when the underlying Riemannian manifold is unknown.
  • Constructing a matrix-valued kernel directly from this connection Laplacian encoding provides a robust inductive bias, capturing the alignment of vector fields with manifold curvature more accurately than independent scalar channels.
  • Incorporating global regularity through the connection Laplacian allows RVGP to perform accurate out-of-sample predictions, super-resolution, and inpainting of vector fields while preserving critical topological features like singularities.

Summary by alphaXiv.

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