Research map / Label-efficient and robust learning / Recognition and long-tail
Loss design and classifier theory
Proposes margin losses, hierarchy-aware objectives and generative versus discriminative classifier analysis.
273 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Recognition and long-tail, keyword cluster “label · classification · classifier”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
58 papers in 2023–24 and 40 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 4.6% to 3.0%.
Open questions the data points to
- Papers report PuzzleMix on CIFAR-100 (Top-1 error (%)) anywhere from 16 to 18.55: comparisons against it depend on whose number you take.
- Papers report ContraGAN on Tiny-ImageNet (FID) anywhere from 24.91 to 28.55: comparisons against it depend on whose number you take.
No longer reported by papers here: CIFAR-10 (accuracy (%), last in 2023); CIFAR-100 (accuracy (%), last in 2024); CIFAR100 (error rate (%), last in 2019); ImageNet (accuracy (%), last in 2024); ImageNet (top-1 accuracy (%), last in 2023).
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 · Towards Open Set Deep Networks (CVPR 2016 · 1,699 citations)
- 2017 · Label-Free Supervision of Neural Networks with Physics and Domain Knowledge (AAAI 2017 · 279 citations)
- 2018 · Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics (CVPR 2018 · 3,007 citations)
- 2019 · CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison (AAAI 2019 · 2,678 citations)
- 2020 · Conditional Gaussian Distribution Learning for Open Set Recognition (CVPR 2020 · 308 citations)
- 2021 · Permuted AdaIN: Reducing the Bias Towards Global Statistics in Image Classification (CVPR 2021 · 89 citations)
- 2022 · Ranking-Based Siamese Visual Tracking (CVPR 2022 · 91 citations)
- 2023 · Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification (CVPR 2023 · 70 citations)
- 2024 · Adaptive Multi-Modal Cross-Entropy Loss for Stereo Matching (CVPR 2024 · 36 citations)
- 2025 · Unlocking Better Closed-Set Alignment Based on Neural Collapse for Open-Set Recognition (AAAI 2025 · 5 citations)
- 2026 · On Revisiting Entropy for Identifying Mislabeled Images (ICML 2026 · 0 citations)
Earliest papers here
- Towards Open Set Deep Networks (CVPR 2016 · 1,699 citations)
- Improved Deep Metric Learning with Multi-class N-pair Loss Objective (NeurIPS 2016 · 1,329 citations)
- CNN-RNN: A Unified Framework for Multi-Label Image Classification (CVPR 2016 · 1,236 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-10 · accuracy (%) (higher is better)
8 papers, 2018–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 92.22 | Maximum-Entropy Fine Grained Classification (NeurIPS 2018 · 49 citations) |
| 2019 | 94.46 | Inter-Class Angular Loss for Convolutional Neural Networks (AAAI 2019 · 12 citations) |
| 2021 | 97.8 | Train a One-Million-Way Instance Classifier for Unsupervised Visual Representation Learning (AAAI 2021 · 3 citations) |
CIFAR-100 · accuracy (%) (higher is better)
8 papers, 2018–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 76.22 | Maximum-Entropy Fine Grained Classification (NeurIPS 2018 · 49 citations) |
| 2019 | 76.74 | Inter-Class Angular Loss for Convolutional Neural Networks (AAAI 2019 · 12 citations) |
| 2021 | 86.2 | Train a One-Million-Way Instance Classifier for Unsupervised Visual Representation Learning (AAAI 2021 · 3 citations) |
| 2024 | 100 | Towards Eliminating Hard Label Constraints in Gradient Inversion Attacks (ICLR 2024 · 0 citations) |
CIFAR100 · Top-1 accuracy (higher is better)
5 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 84.17 | Permuted AdaIN: Reducing the Bias Towards Global Statistics in Image Classification (CVPR 2021 · 89 citations) |
| 2023 | 86.85 | TransHP: Image Classification with Hierarchical Prompting (NeurIPS 2023 · 7 citations) |
CIFAR100 · error rate (%) (lower is better)
5 papers, 2016–2019 · not reported since 2019
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 29.53 | Large-Margin Softmax Loss for Convolutional Neural Networks (ICML 2016 · 431 citations) |
| 2018 | 19.83 | CapProNet: Deep Feature Learning via Orthogonal Projections onto Capsule Subspaces (NeurIPS 2018 · 56 citations) |
ImageNet · accuracy (%) (higher is better)
5 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 80.78 | Consistency Driven Sequential Transformers Attention Model for Partially Observable Scenes (CVPR 2022 · 8 citations) |
| 2024 | 100 | Towards Eliminating Hard Label Constraints in Gradient Inversion Attacks (ICLR 2024 · 0 citations) |
ImageNet · top-1 accuracy (%) (higher is better)
4 papers, 2019–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 61.92 | Positive-Unlabeled Compression on the Cloud (NeurIPS 2019 · 21 citations) |
| 2021 | 79.13 | Permuted AdaIN: Reducing the Bias Towards Global Statistics in Image Classification (CVPR 2021 · 89 citations) |
| 2023 | 82.35 | TransHP: Image Classification with Hierarchical Prompting (NeurIPS 2023 · 7 citations) |
What has been tried
240 papers here report 2,538 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 8%, hurt 19%, 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 |
|---|---|---|---|---|
| Replace uni-modal cross-entropy with multi-modal cross-entropy | SceneFlow test · EPE | 0.84 -> 0.80 | Helped | Adaptive Multi-Modal Cross-Entropy Loss for Stereo Matching (CVPR 2024 · 36 citations) |
| Add the ranking loss LRank | Seq-FaceComp · Fixed-Acc (%) | Helped | Contrastive Learning for DeepFake Classification and Localization via Multi-Label Ranking (CVPR 2024 · 20 citations) | |
| add sorting loss for positive pixels by label certainty | BSDS500 test · AP | .886 -> .895 | Helped | RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses (CVPR 2024 · 20 citations) |
| Replace negative-instance masking with random-value masking | CAMELYON16 bag (slide) · AUC | 0.9015 -> 0.8753 | Hurt | Boosting Multiple Instance Learning Models for Whole Slide Image Classification: A Model-Agnostic Framework Based on Counterfactual Inference (AAAI 2024 · 16 citations) |
| Train with 8 datasets instead of 3 | Test set · DSC | 83.5±13.9 -> 88.7±7.0 | Helped | Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · 8 citations) |
| Remove input-level mixup | CIFAR-100 · Top-1 accuracy (%) | 81.48 (0.46) vs 82.24 (0.04) | No clear effect | Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (AAAI 2024 · 6 citations) |
| use five networks instead of optimal-size ensemble | CIFAR100 symmetric 40% noise · accuracy | 71.3 -> 70.1 | Hurt | United We Stand: Using Epoch-Wise Agreement of Ensembles to Combat Overfit (AAAI 2024 · 6 citations) |
| Add F-DEF loss to DEF | CIFAR10 · AUROC | 94.03 -> 95.69 | Helped | Unlocking Better Closed-Set Alignment Based on Neural Collapse for Open-Set Recognition (AAAI 2025 · 5 citations) |
| Add RMOS training data to MEBOW-Net | MEBOW test · Acc.(5°) | 68.4 -> 70.8 | Helped | Towards Fine-Grained HBOE with Rendered Orientation Set and Laplace Smoothing (AAAI 2024 · 5 citations) |
| add selective refinement during grounding | Dataset 1 · F1 (Flood) | 0.95±0.00 -> 0.96±0.00 | Helped | Spatial-Logic-Aware Weakly Supervised Learning for Flood Mapping on Earth Imagery (AAAI 2024 · 4 citations) |
| Use five neurons per class instead of one | Fashion-MNIST and CIFAR-10, both feature extractors · accuracy | 85.89 ± 0.27 -> 86.74 ± 0.25; 90.61 ± 0.19 -> 91.25 ± 0.20; 61.08 ± 0.17 -> 62.61 ± 0.27; 76.90 ± 0.27 -> 78.29 ± 0.27 | Helped | Neuronal Competition Groups with Supervised STDP for Spike-Based Classification (NeurIPS 2024 · 4 citations) |
| Project image features to remove class proxies | Waterbirds · worst-group accuracy | 81.6 -> 84.3 | Helped | Project-Probe-Aggregate: Efficient Fine-Tuning for Group Robustness (CVPR 2025 · 2 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 |
|---|---|---|---|---|
| PuzzleMix | CIFAR-100 | Top-1 error (%) | 16 – 18.55 | 2 papers, e.g. AAAI 2022 and AAAI 2024 |
| ContraGAN | Tiny-ImageNet | FID | 24.91 – 28.55 | 2 papers, e.g. NeurIPS 2021 and NeurIPS 2021 |
| PuzzleMix | Tiny-ImageNet | Top-1 error (%) | 32.98 – 35.79 | 2 papers, e.g. AAAI 2024 and AAAI 2022 |
What reviewers will expect
Datasets and baselines named in the 23 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-100 · 7 of 23 papers
- CIFAR10 · 7 of 23 papers
- ImageNet · 5 of 23 papers
- Tiny-ImageNet · 3 of 23 papers
- ImageNet-1k · 3 of 23 papers
- MNIST · 3 of 23 papers
- Caltech101 · 3 of 23 papers
- Food101 · 3 of 23 papers
Baselines
- MixUp · 3 of 23 papers
- CLIP · 2 of 23 papers
- CoOp · 2 of 23 papers
- MaPLe · 2 of 23 papers
- KgCoOp · 2 of 23 papers
- DePT · 2 of 23 papers
- label smoothing · 2 of 23 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 98 papers; from OpenAlex and the papers' affiliations).
- Zhejiang University · 4 papers
- Wuhan University · 3 papers
- Tsinghua University · 3 papers
- National University of Singapore · 3 papers
- Chinese Academy of Sciences · 3 papers
- Nanyang Technological University · 2 papers
- Baidu · 2 papers
- Peng Cheng Laboratory · 2 papers
- Auburn University · 2 papers
- Huazhong University of Science and Technology · 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
- Partial and multi-label learning (345 papers)
- Group-robust classification (251 papers)
- Learning from unlabeled and positive-unlabeled data (240 papers)
- Confidence calibration objectives (129 papers)
- Deep active learning (76 papers)
Elsewhere in the topic
- Multilingual language representation pretraining (507 papers)
- Contrastive representation theory (495 papers)
- Dense features and foundation priors (482 papers)
- Correspondence and causal representation learning (447 papers)
- Data augmentation and generalization bias (409 papers)
- Recommendation representation learning (207 papers)
- Margin-based face recognition losses (124 papers)
- Supervised person re-identification (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).