Reproduce “Emergent Riemannian geometry over learning discrete computations on continuous 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

Julian Brandon, Angus Chadwick, Arthur Pellegrino

The authors use pullback metrics to interpret how networks map changing inputs to discrete outputs. The metric measures activation-change dot products as task coordinates change. In toy XOR networks, hidden layers mark input transitions and output layers reflect the Boolean operation. Rich-learning networks generalize to held-out points, unlike lazy-learning networks, but findings remain limited to small, hand-crafted settings.

Key insights

  • Neural network computation for discrete tasks on continuous manifolds can be functionally decomposed: early layers discretize continuous variables, and later layers perform logical operations on these binarized features.
  • Effective generalization in neural networks is strongly linked to the emergence of structured, low-dimensional representational geometries with specific curvature patterns in the rich learning regime.
  • The representational geometry adapts to input noise during training, with changes in metric values and curvature reflecting the network's uncertainty and approximating a Bayesian posterior distribution.

Summary by alphaXiv.

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