Reproduce “Disentangling by Subspace Diffusion” 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

David Pfau, Irina Higgins, Aleksandar Botev, Sébastien Racanière

DeepMind researchers developed GEOMANCER, a nonparametric algorithm for unsupervised, symmetry-based disentangling of data manifolds, leveraging differential geometry to factorize Lie groups solely from observed object orbits. The method successfully disentangled complex product manifolds with up to five submanifolds, including 3D rotations, when provided with accurate metric information.

Key insights

  • The problem of unsupervised disentanglement can be reduced to the problem of unsupervised metric learning, as GEOMANCER's performance is critically dependent on the accuracy of local metric information.
  • The properties of infinitesimal transformations and the "failures" of the parallelogram model (cycle-inconsistency) can serve as a learning signal for disentanglement, formalized by holonomy theory.
  • The approach successfully handles transformations with non-trivial holonomy, such as 3D rotations, which were largely unaddressed by previous symmetry-based disentanglement methods.

Summary by alphaXiv.

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