Research map / Learning theory and algorithms / Kernels and regression theory
Empirical generalization measures
Measure and predict generalization of deep networks through experiments and scaling laws.
61 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Kernels and regression theory, keyword cluster “generalization bounds · pac · generalization error”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
14 papers in 2023–24 and 16 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 5.6% to 6.5%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Exploring Generalization in Deep Learning (NeurIPS 2017 · 450 citations)
- 2018 · Sensitivity and Generalization in Neural Networks: an Empirical Study (ICLR 2018 · 285 citations)
- 2019 · The role of over-parametrization in generalization of neural networks (ICLR 2019 · 144 citations)
- 2020 · Fantastic Generalization Measures and Where to Find Them (ICLR 2020 · 147 citations)
- 2021 · In Search of Lost Domain Generalization (ICLR 2021 · 79 citations)
- 2022 · Why Robust Generalization in Deep Learning is Difficult: Perspective of Expressive Power (NeurIPS 2022 · 6 citations)
- 2023 · Generalization and Estimation Error Bounds for Model-based Neural Networks (ICLR 2023 · 3 citations)
- 2024 · Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains (AAAI 2024 · 5 citations)
- 2025 · Generalization Analysis for Deep Contrastive Representation Learning (AAAI 2025 · 1 citation)
- 2026 · The Data Manifold under the Microscope (ICML 2026 · 0 citations)
Earliest papers here
- Exploring Generalization in Deep Learning (NeurIPS 2017 · 450 citations)
- Sensitivity and Generalization in Neural Networks: an Empirical Study (ICLR 2018 · 285 citations)
- Stronger Generalization Bounds for Deep Nets via a Compression Approach (ICML 2018 · 81 citations)
What has been tried
45 papers here report 359 measured changes (a component added, removed or swapped, compared with a baseline): helped 62%, no clear effect 11%, hurt 21%, mixed 7%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use MUNIT or VGG11 architecture instead of the original architecture | ColoredMNIST · test accuracy | Helped | Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains (AAAI 2024 · 5 citations) | |
| Augment the loss with an output-representation norm constraint | DCRL unsupervised-risk generalization bound · spectral-norm dependence | product of squared spectral norms -> single product of spectral norms | Helped | Generalization Analysis for Deep Contrastive Representation Learning (AAAI 2025 · 1 citation) |
| Reduce second-layer key and query embeddings with SVD or clustering | Dyck, unseen structures · prediction similarity to the original model | Hurt | Interpretability Illusions in the Generalization of Simplified Models (ICML 2024 · 1 citation) | |
| Increase the counting subspace dimension from 2D to 3D | CIFAR-10, ResNet18 · correlation with generalization gap | 0.96 -> 0.97 | No clear effect | Understanding Nonlinear Implicit Bias via Region Counts in Input Space (ICML 2025 · 0 citations) |
| replace ReLU activation with tanh | Streaming parity · infinite generalization | No clear effect | Algorithm Development in Neural Networks: Insights from the Streaming Parity Task (ICML 2025 · 0 citations) | |
| Add an intermediate fine-tuning step through subregime 2A before the final regime-1 LR | CIFAR-10 · test accuracy | Helped | Where Do Large Learning Rates Lead Us? (NeurIPS 2024 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 12 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 4 of 12 papers
- MNIST · 3 of 12 papers
- CIFAR-100 · 2 of 12 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 30 papers; from OpenAlex and the papers' affiliations).
- The University of Hong Kong · 3 papers
- Meta · 3 papers
- University of Oxford · 3 papers
- Singapore Management University · 2 papers
- University College London · 2 papers
- Stanford University · 2 papers
- Google · 2 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Statistical perspectives on generalization (99 papers)
- Information-theoretic generalization bounds (94 papers)
- Generalization theory of deep networks (67 papers)
- PAC-Bayes bounds (39 papers)
Elsewhere in the topic
- RKHS kernel methods (127 papers)
- Neural tangent kernel analysis (109 papers)
- Kernel regression under shift (89 papers)
- High-dimensional asymptotic generalization (82 papers)
- High-dimensional linear regression theory (61 papers)
- Kernel ridge regression analysis (57 papers)
- Double descent and overfitting (53 papers)
- Random feature approximation (36 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).