Flattening vs. Tangling: Negative-Eigenvalue Geodesic Curvature Separates Generalization from Memorization Where Linear Effective Rank Fails
Geometric and empirical demonstration that negative-eigenvalue geodesic curvature g3 on connectivity-guaranteed MST+kNN hidden-activation graphs cleanly separates neural network generalization (monotonic manifold flattening, -43% to -73%) from random-label memorization (active manifold tangling, +11% to +75%) across ResNet-CNNs and 5-seed MLPs (>7 sigma separation, robust across neighborhood scales k) where linear covariance rank and Shannon effective rank fail or invert by up to 2.8x. Direct audits of memorized training clouds uncover intermediate-layer training tangling and a label-free readout Train-Test curvature gap (Delta g3 > 0.54), alongside two-stage asymmetric flattening in grokking.
Authors
- Shoji Fujita
Publication Details
- Journal
- Figshare
- Published
- 2026-11-20
- DOI
- https://doi.org/10.6084/m9.figshare.35074318
- Primary Topic
- Neural Networks and Applications
- Type
- preprint