Research map / Label-efficient and robust learning / Recognition and long-tail
Gauge-invariant representation holonomy
Vasileios Sevetlidis, George Pavlidis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use separate k-NN sets per edge endpoint instead of shared midpoint neighbors (ablation) | holonomy estimator · MNIST Hidden1, r=0.10 | h_norm: 6.42e-07 -> 0.222 | Hurt |
| add shared-midpoint neighbor selection (ablation) | holonomy estimator · MNIST Hidden1 | h_norm bias: +2.22e-1 avoided | Improved |
| allow O(p) reflections instead of SO(p) (ablation) | holonomy estimator · MNIST Hidden1, r=0.10 | h_norm: text +5.37e-7 vs appendix table 6.42e-07 -> 6.42e-07 | Mixed |
| per-neighborhood whitening instead of global (ablation) | holonomy estimator · MNIST Hidden1, r=0.10 | h_norm: text +1.59e-7 vs appendix 6.42e-07 -> 2.22e-07 | Mixed |
| vary (k,q) over {96,128,192} x {32,64,96} (ablation) | holonomy estimator · MNIST Hidden1, r=0.10 | h_norm: change 7.20e-7 (SD 2.86e-7) | No clear effect |
| increase standardization pool from 10^3 to 8x10^3 (ablation) | holonomy estimator · MNIST Hidden1, r=0.10 | h_norm: 4.05e-6 -> 4.06e-6 | No clear effect |
| random plane instead of local PCA plane (ablation) | holonomy estimator · MNIST Hidden1, r=0.10 | h_norm: -1.92e-8 | No clear effect |
| holonomy correlates with robustness at regime level | holonomy vs FGSM accuracy · CIFAR-10 layer2, 4 regimes | FGSM accuracy: r approx 0.94 | Improved |
| condition on clean accuracy | holonomy vs robustness · CIFAR-10 layer2, per seed | partial correlation: FGSM 0.223, PGD 0.276, corruption 0.027 | No clear effect |
| replace nonlinearities with identity (linear null) (ablation) | MLP · MNIST | h_norm: to 9.57e-9 noise level | Improved |
Closest papers by meaning
- Connecting Neural Models Latent Geometries with Relative Geodesic Representations (NeurIPS 2025 · similarity 0.85)
- Parameter symmetries determine representational geometry in overparameterized nonlinear networks (NeurIPS 2026 · similarity 0.85)
- From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication (ICLR 2024 · similarity 0.84)
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.84)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.84)
- Intrinsic Dimension Correlation: uncovering nonlinear connections in multimodal representations (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
