Research map / Label-efficient and robust learning / Noisy labels
Loss correction and noise robustness theory
Corrects losses and analyzes how overfitting and subclass structure affect noisy training.
77 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Noisy labels, keyword cluster “label noise · transition · instance”
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.
17 papers in 2023–24 and 24 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 14.0% to 19.7%.
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 · Data Programming: Creating Large Training Sets, Quickly (NeurIPS 2016 · 356 citations)
- 2017 · Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach (CVPR 2017 · 1,463 citations)
- 2018 · Reinforcement Learning for Relation Classification From Noisy Data (AAAI 2018 · 302 citations)
- 2019 · Label-Noise Robust Generative Adversarial Networks (CVPR 2019 · 68 citations)
- 2020 · Transfer Learning From Synthetic to Real-Noise Denoising With Adaptive Instance Normalization (CVPR 2020 · 225 citations)
- 2021 · Unlearnable Examples: Making Personal Data Unexploitable (ICLR 2021 · 45 citations)
- 2022 · Robust Training under Label Noise by Over-parameterization (ICML 2022 · 26 citations)
- 2023 · Free Lunch for Domain Adversarial Training: Environment Label Smoothing (ICLR 2023 · 24 citations)
- 2024 · Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks (AAAI 2024 · 8 citations)
- 2025 · The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation (CVPR 2025 · 2 citations)
- 2026 · Data Augmentation of Contrastive Learning is Estimating Positive-incentive Noise (ICML 2026 · 17 citations)
Earliest papers here
- Data Programming: Creating Large Training Sets, Quickly (NeurIPS 2016 · 356 citations)
- Training Deep Neural Networks via Direct Loss Minimization (ICML 2016 · 69 citations)
- Re-Active Learning: Active Learning with Relabeling (AAAI 2016 · 60 citations)
What has been tried
60 papers here report 623 measured changes (a component added, removed or swapped, compared with a baseline): helped 63%, no clear effect 10%, hurt 23%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add class-level curriculum learning (C-CL) to a model without curriculum learning | CIFAR-100N, Food101, and miniWV, 5-way 5-shot · accuracy | Helped | Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks (AAAI 2024 · 8 citations) | |
| Remove scaling noise (use shift-only GNI) | CIFAR-100 validation · accuracy | Hurt | Ghost Noise for Regularizing Deep Neural Networks (AAAI 2024 · 2 citations) | |
| Use SGD-Shapley labels, which have non-negligible bias, instead of unbiased Shapley estimators | ImageNette · Attribution estimation error | Hurt | Stochastic Amortization: A Unified Approach to Accelerate Feature and Data Attribution (NeurIPS 2024 · 2 citations) | |
| Add batch normalization and data augmentation to Label Wave | CIFAR-10, 40% symmetric noise · test accuracy | 63.00±0.86% -> 66.79±0.39% | Helped | Early Stopping Against Label Noise Without Validation Data (ICLR 2024 · 1 citation) |
| add a Task Encoder rather than a fully connected network | Handwriting I · accuracy | 93.6% -> 96.1% | Helped | DeCorrNet: Enhancing Neural Decoding Performance by Eliminating Correlations in Noise (AAAI 2025 · 0 citations) |
| Remove a portion of examples with high influence scores | MNIST, CIFAR-10, and CIFAR-100 · test accuracy | Hurt | Rethinking Benign Overfitting in Two-Layer Neural Networks (ICML 2025 · 0 citations) | |
| Use hard rather than soft pseudo-labels | Noisy Gaussian mixture · generalization error improvement | No clear effect | The Effect of Optimal Self-Distillation in Noisy Gaussian Mixture Model (NeurIPS 2025 · 0 citations) | |
| Add kNN label spreading before RAD | CelebA at 20% and 30% label noise · WGA | 0 (0) -> 82.29 (1.25); 0 (0) -> 79.5 (1.91) | Helped | Enhancing Robustness of Last Layer Two-Stage Fair Model Corrections (NeurIPS 2024 · 0 citations) |
| Remove the consistency regularization term (α = 0) | Artificially noised confidence-difference datasets · accuracy | Hurt | Confidence Difference Reflects Various Supervised Signals in Confidence-Difference Classification (ICML 2025) | |
| add Matching loss | LF-WikiSeeAlsoTitles-320K · P@1 | 34.12 -> 34.62 | Helped | MOGIC: Metadata-infused Oracle Guidance for Improved Extreme Classification (ICML 2025) |
What reviewers will expect
Datasets and baselines named in the 16 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 · 5 of 16 papers
- MNIST · 3 of 16 papers
- CIFAR-100 · 2 of 16 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 41 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 3 papers
- The University of Sydney · 3 papers
- Google · 3 papers
- Stanford University · 2 papers
- University of Washington · 2 papers
- University of California, San Diego · 2 papers
- Shandong University · 2 papers
- Korea Advanced Institute of Science and Technology · 2 papers
- Princeton University · 2 papers
Use it
Download BibTeX of 39 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
- Instance-dependent noise modeling (106 papers)
- Crowdsourced label aggregation (23 papers)
Elsewhere in the topic
- Sample selection and co-training (86 papers)
- Noise-aware loss and pairwise methods (66 papers)
- Contrastive learning on noisy data (55 papers)
- Noisy cross-modal hashing and multi-label (49 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).