Reproduce “Neural Feature Geometry Evolves as Discrete Ricci Flow” 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
Moritz Hehl, Max von Renesse, Melanie Weber
Researchers from Leipzig University and Harvard University developed a framework connecting the evolution of neural network feature geometry to discrete Ricci flow. They theoretically demonstrate the essential role of non-linear activations in reshaping feature spaces and empirically show that learning dynamics consistently align with Ricci flow, leading to improved class separability and offering novel geometry-informed heuristics for early-stopping and optimal network depth selection.
Key insights
- Non-linear activation functions are indispensable for genuine geometric transformations, as linear networks largely preserve initial feature geometries and graph structures.
- Neural network learning consistently drives feature geometry to evolve in a manner analogous to discrete Ricci flow, characterized by predominantly negative local Ricci evolution coefficients.
- This geometric evolution enhances class separability by strengthening community structures within the feature graphs, aligning learned features with underlying class labels.
- Monitoring geometric evolution metrics can provide practical guidance for deep learning, such as identifying a 'critical depth' for optimal network architecture and a reliable early-stopping criterion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.