Research map / Optimization / Implicit bias and generalization
Generalization Below the Edge of Stability: The Role of Data Geometry
Tongtong Liang, Alex Cloninger, Rahul Parhi, Yu-Xiang Wang
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 |
|---|---|---|---|
| change the input distribution from the uniform sphere to a union of 20 lines | two-layer ReLU network trained with GD · Synthetic quadratic teacher, d=50 | final true MSE: 1.0249 -> 0.07 | Improved |
| concentrate input mass away from the boundary (larger Beta-radial alpha, 0.1 to 5.0) | two-layer ReLU network, d=5 · Beta-radial synthetic data | log-log error slope vs n | Improved |
| increase ambient dimension d from 10 to 500 for a union of 20 lines | two-layer ReLU network · union of 20 lines (m=1) | regression slope of error vs n | No clear effect |
| replace Gaussian inputs with MNIST inputs | two-layer ReLU network, n=30000 · Gaussian vs MNIST inputs (ground-truth teacher) | time to memorize noisy labels | Improved |
| add low-dimensional structure to the data (sphere to mixture) | two-layer ReLU network · neuron data activation rate | share of neurons firing on 10-40% of data | Improved |
| move data mass toward the sphere surface (shallow half-space depth) | two-layer ReLU network · MNIST samples | prediction error vs half-space depth | Hurt |
Closest papers by meaning
- Stable Minima of ReLU Neural Networks Suffer from the Curse of Dimensionality: The Neural Shattering Phenomenon (NeurIPS 2025 · similarity 0.91)
- Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.90)
- Neural Characteristic Activation Analysis and Geometric Parameterization for ReLU Networks (NeurIPS 2024 · similarity 0.89)
- Stable Minima Cannot Overfit in Univariate ReLU Networks: Generalization by Large Step Sizes (NeurIPS 2024 · similarity 0.88)
- Bias of Stochastic Gradient Descent or the Architecture: Disentangling the Effects of Overparameterization of Neural Networks (ICML 2024 · similarity 0.88)
- Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking (NeurIPS 2025 · similarity 0.88)
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.
