Research map / Learning theory and algorithms / Kernels and regression theory
In search of robust measures of generalization
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 |
|---|---|---|---|
| Evaluate measures by worst-case sign-error across environments instead of averaging sign-error | Generalization measures on CIFAR-10 and SVHN · All environments | robust sign-error | Mixed |
| Vary training-set size | Most generalization measures · Train-size environments | sign-error | Helped |
| Vary network width | Generalization measures · Width environments | robust sign-error | Hurt |
| Vary network depth | Generalization measures · Depth environments | robust sign-error | Hurt |
| Use norm-based measures to predict changes across learning rates | Generalization measures · Learning-rate environments | sign-error | Helped |
| Use PAC-Bayes-based measures | Generalization measures · Depth environments | sign-error | Mixed |
| Restrict evaluation to SVHN | Generalization measures · SVHN depth environments | robust sign-error | Helped |
| Include ranking pairs with small generalization-gap differences instead of filtering Monte Carlo noise (ablation) | Generalization measures · Coupled-network ranking experiments | mean sign-error | Hurt |
| Use the exact FFT-based spectral-norm calculation rather than the approximation (ablation) | Spectral generalization measures · CIFAR-10 and SVHN experiments | ranking performance | No clear effect |
Closest papers by meaning
- Fantastic Generalization Measures and Where to Find Them (ICLR 2020 · similarity 0.92)
- Generalizability of Neural Networks Minimizing Empirical Risk Based on Expressive Power (ICLR 2025 · similarity 0.92)
- Measuring Generalization with Optimal Transport (NeurIPS 2021 · similarity 0.92)
- Fantastic Generalization Measures are Nowhere to be Found (ICLR 2024 · similarity 0.91)
- Understanding Generalization from Embedding Dimension and Distributional Convergence (ICML 2026 · similarity 0.90)
- Why Robust Generalization in Deep Learning is Difficult: Perspective of Expressive Power (NeurIPS 2022 · similarity 0.90)
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.