Research map / Label-efficient and robust learning / Recognition and long-tail
Data augmentation and generalization bias
Studies augmentation, data properties and influence-based fixes for generalization and dataset bias.
409 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Recognition and long-tail, keyword cluster “representations · transfer · unsupervised”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 5 tracked benchmarks.
124 papers in 2023–24 and 119 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 9.9% to 9.1%.
Open questions the data points to
- Waterbirds (worst-group accuracy (%)): best went from 89.67 (up to 2024) to 92.2 in 2025 — NeuronTune: Towards Self-Guided Spurious Bias Mitigation (ICML 2025 · 0 citations)
- Papers report ERM on Waterbirds (worst-group accuracy) anywhere from 57.9 to 74.81: comparisons against it depend on whose number you take.
- Papers report JTT on Waterbirds (worst-group accuracy (%)) anywhere from 63.6 to 86.7: comparisons against it depend on whose number you take.
What looks settled
- CIFAR-100 (accuracy): best up to 2024 89.22, best in 2025–26 79.5 from 2 papers, at most 0.2% better, if at all.
- CIFAR-10 (accuracy (%)): best up to 2024 98.7, best in 2025–26 96 from 2 papers, at most 0.2% better, if at all.
- MultiNLI (worst-group accuracy (%)): best up to 2024 76, best in 2025–26 75.88 from 2 papers, at most 0.2% better, if at all.
No longer reported by papers here: ImageNet-C (mCE (%), last in 2024).
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
- 2016 · Transfer Learning from Deep Features for Remote Sensing and Poverty Mapping (AAAI 2016 · 428 citations)
- 2017 · Generative Models and Model Criticism via Optimized Maximum Mean Discrepancy (ICLR 2017 · 163 citations)
- 2018 · On the importance of single directions for generalization (ICLR 2018 · 235 citations)
- 2019 · SpotTune: Transfer Learning Through Adaptive Fine-Tuning (CVPR 2019 · 472 citations)
- 2020 · Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigation (CVPR 2020 · 273 citations)
- 2021 · Deep Stable Learning for Out-of-Distribution Generalization (CVPR 2021 · 283 citations)
- 2022 · Why do tree-based models still outperform deep learning on typical tabular data? (NeurIPS 2022 · 679 citations)
- 2023 · Improved Distribution Matching for Dataset Condensation (CVPR 2023 · 77 citations)
- 2024 · Preserving Fairness Generalization in Deepfake Detection (CVPR 2024 · 94 citations)
- 2025 · Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · 23 citations)
- 2026 · Inequality in the Age of Pseudonymity (AAAI 2026 · 1 citation)
Earliest papers here
- Transfer Learning from Deep Features for Remote Sensing and Poverty Mapping (AAAI 2016 · 428 citations)
- Unsupervised Feature Selection with Structured Graph Optimization (AAAI 2016 · 323 citations)
- Unsupervised Feature Selection on Networks: A Generative View (AAAI 2016 · 53 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
CIFAR-100 · accuracy (higher is better)
5 papers, 2020–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 89.22 | Do Adversarially Robust ImageNet Models Transfer Better? (NeurIPS 2020 · 110 citations) |
CIFAR-10 · accuracy (%) (higher is better)
4 papers, 2020–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 98.7 | Do Adversarially Robust ImageNet Models Transfer Better? (NeurIPS 2020 · 110 citations) |
ImageNet-C · mCE (%) (lower is better)
4 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 52.1 | VITA: A Multi-Source Vicinal Transfer Augmentation Method for Out-of-Distribution Generalization (AAAI 2022 · 4 citations) |
MultiNLI · worst-group accuracy (%) (higher is better)
4 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 76 | Improving Subgroup Robustness via Data Selection (NeurIPS 2024 · 0 citations) |
Waterbirds · worst-group accuracy (%) (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 89.67 | A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · 0 citations) |
| 2025 | 92.2 | NeuronTune: Towards Self-Guided Spurious Bias Mitigation (ICML 2025 · 0 citations) |
What has been tried
330 papers here report 3,419 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 8%, hurt 19%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add reconstruction loss | Celeb-DF · FFPR | 5.47% | Helped | Preserving Fairness Generalization in Deepfake Detection (CVPR 2024 · 94 citations) |
| add 2nd-order moment regularization to DM | CIFAR-10, IPC=10 · test accuracy | 48.9 -> 52.1 | Helped | M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy (AAAI 2024 · 27 citations) |
| Add token shuffling and token mixing branches | CDF · video-level AUC (%) | 90.70% to 93.13% | Helped | Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · 23 citations) |
| Add Fo-Mixup augmentation | Celeb-DF-v2 · AUC | 69.2 -> 74.8 | Helped | FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 2025 · 18 citations) |
| Use AFA in the auxiliary component rather than only in the main component | C10 · RA | 83.25 -> 88.22 | Helped | Fourier-basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification (CVPR 2024 · 18 citations) |
| Add class centralization to DM | CIFAR10 · accuracy | 48.5±0.2 -> 52.2±0.3 | Helped | Exploiting Inter-sample and Inter-feature Relations in Dataset Distillation (CVPR 2024 · 17 citations) |
| Add prime reinforcement regularization | unbiased test sets across the evaluated datasets · accuracy | Helped | Navigate Beyond Shortcuts: Debiased Learning Through the Lens of Neural Collapse (CVPR 2024 · 10 citations) | |
| add context labels to vanilla attention values | CA · RMSE | 0.484 -> 0.489 | Hurt | TabR: Tabular Deep Learning Meets Nearest Neighbors (ICLR 2024 · 9 citations) |
| Add random cropping to no augmentation | YCD, 10K training images per dataset · accuracy | 43.2 -> 66.1 | Helped | A Decade's Battle on Dataset Bias: Are We There Yet? (ICLR 2025 · 8 citations) |
| Add feature similarity and CDS constraint using L2 norm | CIFAR-10 test, 0.5% sampling · accuracy | 34.0 -> 32.8; 1.2% | Hurt | Contributing Dimension Structure of Deep Feature for Coreset Selection (AAAI 2024 · 7 citations) |
| add the gating-model loss L1 | WebKB · accuracy | 23.22 -> 41.11 | Helped | GraphMETRO: Mitigating Complex Graph Distribution Shifts via Mixture of Aligned Experts (NeurIPS 2024 · 7 citations) |
| Use nonlinear hidden activations instead of linear activations | Synthetic vector classification · shortcut bias | Helped | On the Foundations of Shortcut Learning (ICLR 2024 · 6 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| ERM | Waterbirds | worst-group accuracy | 57.9 – 74.81 | 3 papers, e.g. NeurIPS 2024 and ICLR 2025 |
| JTT | Waterbirds | worst-group accuracy (%) | 63.6 – 86.7 | 2 papers, e.g. NeurIPS 2024 and ICLR 2025 |
| ERM | CelebA | worst-group accuracy | 47.78 – 57.6 | 2 papers, e.g. NeurIPS 2024 and ICML 2023 |
| IDM | CIFAR100, IPC=50 | accuracy | 46.5 – 49.9 | 2 papers, e.g. NeurIPS 2025 and CVPR 2024 |
What reviewers will expect
Datasets and baselines named in the 76 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 · 19 of 76 papers
- CIFAR-100 · 12 of 76 papers
- ImageNet-1K · 12 of 76 papers
- CelebA · 11 of 76 papers
- Tiny-ImageNet · 7 of 76 papers
- MNIST · 6 of 76 papers
- MultiNLI · 6 of 76 papers
- ImageNet-100 · 5 of 76 papers
Baselines
- ERM · 8 of 76 papers
- JTT · 7 of 76 papers
- DFR · 6 of 76 papers
- Random · 5 of 76 papers
- LfF · 4 of 76 papers
- RDED · 4 of 76 papers
- RankSim · 3 of 76 papers
- ConR · 3 of 76 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 243 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 8 papers
- Google · 8 papers
- Zhejiang University · 8 papers
- Amazon · 8 papers
- Nanjing University · 7 papers
- New York University · 7 papers
- Stanford University · 7 papers
- Carnegie Mellon University · 6 papers
- Harvard University · 6 papers
- Tencent · 5 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
- Contrastive representation theory (495 papers)
- Correspondence and causal representation learning (447 papers)
- Dense features and foundation priors (482 papers)
- Multilingual language representation pretraining (507 papers)
- Recommendation representation learning (207 papers)
Elsewhere in the topic
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 papers)
- Learning from unlabeled and positive-unlabeled data (240 papers)
- Confidence calibration objectives (129 papers)
- Margin-based face recognition losses (124 papers)
- Supervised person re-identification (108 papers)
- Self-explaining and visual explanation networks (108 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).