Research map / Optimization / Implicit bias and generalization
Finding Generalization Measures by Contrasting Signal and Noise
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add noise dataset term (REF Complexity vs. REF w/o noi) (ablation) | REF Complexity measure · CIFAR-10 and CIFAR-100 over 100 experiments | correlation with test accuracy: -0.988 vs -0.585 (CIFAR-10); -0.960 vs -0.481 (CIFAR-100) | Helped |
| Remove initial loss normalization (REF w/o init) (ablation) | REF Complexity measure · CIFAR-10 over 100 experiments | correlation with test accuracy: -0.988 vs -0.988 (paper states initial loss does not affect much) | No clear effect |
| Train longer (epoch T from 50 to 500) (ablation) | REF Complexity · CIFAR-10 | correlation with test accuracy: -0.9743 at T=50 to -0.2202 at T=500 | Hurt |
| Use stochastic (small batch) instead of deterministic training in REF measure | SGD vs GD in REF Complexity · CIFAR-10 Table 1 (batch size variation) | correlation with test accuracy | Helped |
| Use REF Complexity vs. existing norm/sharpness baselines (e.g., PATH-NORM, PB-M-FLATNESS) | generalization measure · CIFAR-10 Table 1 across batch size, LR, dropout | consistency of negative correlation with test accuracy: REF -0.997/-0.998/-0.965 vs baselines with sign flips (e.g., PB-M-FLATNESS +0.977 on LR) | Helped |
Closest papers by meaning
- In search of robust measures of generalization (NeurIPS 2020 · similarity 0.89)
- Measuring Generalization with Optimal Transport (NeurIPS 2021 · similarity 0.89)
- Generalizability of Neural Networks Minimizing Empirical Risk Based on Expressive Power (ICLR 2025 · similarity 0.88)
- Model-agnostic Measure of Generalization Difficulty (ICML 2023 · similarity 0.88)
- Fantastic Generalization Measures are Nowhere to be Found (ICLR 2024 · similarity 0.88)
- Towards Generalization beyond Pointwise Learning: A Unified Information-theoretic Perspective (ICML 2024 · similarity 0.87)
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.