Research map / Label-efficient and robust learning / Recognition and long-tail
Prototype and concept dictionary models
Uses prototypical parts and sparse concept dictionaries for interpretable predictions.
104 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Recognition and long-tail, keyword cluster “explanations · concept · interpretable”
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.
27 papers in 2023–24 and 42 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 2.2% to 3.2%.
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 · Towards Deep Interpretability (MUS-ROVER II): Learning Hierarchical Representations of Tonal Music (ICLR 2017 · 3 citations)
- 2018 · Towards Black-box Iterative Machine Teaching (ICML 2018 · 19 citations)
- 2019 · Knockoff Nets: Stealing Functionality of Black-Box Models (CVPR 2019 · 537 citations)
- 2020 · AdaCare: Explainable Clinical Health Status Representation Learning via Scale-Adaptive Feature Extraction and Recalibration (AAAI 2020 · 112 citations)
- 2021 · XProtoNet: Diagnosis in Chest Radiography With Global and Local Explanations (CVPR 2021 · 124 citations)
- 2022 · How Faithful is your Synthetic Data? Sample-level Metrics for Evaluating and Auditing Generative Models (ICML 2022 · 22 citations)
- 2023 · Disentangled CVAEs with Contrastive Learning for Explainable Recommendation (AAAI 2023 · 14 citations)
- 2024 · SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · 15 citations)
- 2025 · Learning Concept Prerequisite Relation via Global Knowledge Relation Optimization (AAAI 2025 · 4 citations)
- 2026 · Explainable Melanoma Diagnosis with Contrastive Learning and LLM-based Report Generation (AAAI 2026 · 1 citation)
Earliest papers here
- Towards Deep Interpretability (MUS-ROVER II): Learning Hierarchical Representations of Tonal Music (ICLR 2017 · 3 citations)
- Towards Black-box Iterative Machine Teaching (ICML 2018 · 19 citations)
- Learning from discriminative feature feedback (NeurIPS 2018 · 8 citations)
What has been tried
73 papers here report 695 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 11%, hurt 18%, 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 differentiable PAG Top-K patch selection | TCGA-Lung · accuracy | 0.859 -> 0.884 | Helped | SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · 15 citations) |
| Increase SpLiCE decomposition sparsity to 512 | CIFAR100 · zero-shot accuracy | 0.750 -> 0.768 | Helped | Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE) (NeurIPS 2024 · 13 citations) |
| Finetune all modules during Stage 2 | In-the-wild and MLAAD-EN · EER | 12.9 -> 12.5; 13.5 -> 10.7 | Helped | SLIM: Style-Linguistics Mismatch Model for Generalized Audio Deepfake Detection (NeurIPS 2024 · 6 citations) |
| add document-concept containment task | LectureBank · ACC | 0.8171 -> 0.8232 | Helped | Learning Concept Prerequisite Relation via Global Knowledge Relation Optimization (AAAI 2025 · 4 citations) |
| add the disparity score | SEA, Random RBF, Sine, Electricity, Weather, and Spam · accuracy | without disparity -> Quilt: SEA .890 -> .888; Random RBF .828 -> .833; Sine .890 -> .936; Electricity .711 -> .728; Weat | Mixed | Quilt: Robust Data Segment Selection against Concept Drifts (AAAI 2024 · 4 citations) |
| restrict CBC to positive reasoning only | CUB test · accuracy | 87.8±0.1 -> 28.6±0.8 | Hurt | A Robust Prototype-Based Network with Interpretable RBF Classifier Foundations (AAAI 2025 · 3 citations) |
| Use two concept-convolution windows rather than one | spatiotemporal event forecasting datasets · prediction performance | Helped | GeoPro-Net: Learning Interpretable Spatiotemporal Prediction Models Through Statistically-Guided Geo-Prototyping (AAAI 2025 · 3 citations) | |
| add random data augmentation | Evidence Inference · Token F1 | 26.8 ± 0.2 -> 33.1 ± 0.2 | Helped | Towards Faithful Explanations: Boosting Rationalization with Shortcuts Discovery (ICLR 2024 · 2 citations) |
| add concept loss | COVID-CT · accuracy | 81.36 -> 84.75 | Helped | Towards Multi-dimensional Explanation Alignment for Medical Classification (NeurIPS 2024 · 2 citations) |
| Relax the convex-hull constraint with δ=1.0 rather than δ=0 | ConvNeXt, 32k dictionary · plausibility score | 0.3999 -> 0.5581 | Helped | Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models (ICML 2025 · 1 citation) |
| Add CIG to CIDR | SST2, IMDB, and Rotten Tomatoes · FMS | 0.459 -> 0.592 | Helped | CIDR: A Cooperative Integrated Dynamic Refining Method for Minimal Feature Removal Problem (AAAI 2024 · 1 citation) |
| use global-to-local feature selection instead of global-only selection | colon · balanced accuracy (%) | 75.95 -> 83.95 | Helped | ProtoGate: Prototype-based Neural Networks with Global-to-local Feature Selection for Tabular Biomedical Data (ICML 2024 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 25 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MNIST · 4 of 25 papers
- CUB · 3 of 25 papers
- ImageNet · 3 of 25 papers
- ImageNet1K · 3 of 25 papers
- CIFAR-10 · 2 of 25 papers
- CIFAR-100 · 2 of 25 papers
- Stanford Cars · 2 of 25 papers
Baselines
- PIP-Net · 3 of 25 papers
- ProtoPool · 3 of 25 papers
- NB · 2 of 25 papers
- SVM · 2 of 25 papers
- TopK SAE · 2 of 25 papers
- JumpReLU SAE · 2 of 25 papers
- KMeans · 2 of 25 papers
- Q-SENN · 2 of 25 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 69 papers; from OpenAlex and the papers' affiliations).
- Peking University · 3 papers
- Harvard University · 3 papers
- TU Berlin · 3 papers
- East China Normal University · 2 papers
- University of Iowa · 2 papers
- Massachusetts Institute of Technology · 2 papers
- Stanford University · 2 papers
- University of Pennsylvania · 2 papers
- Technische Universität Berlin · 2 papers
Use it
Download BibTeX of 38 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
- Self-explaining and visual explanation networks (108 papers)
- Post-hoc explanation analysis (86 papers)
- Feature attribution methods (51 papers)
- Concept bottleneck models (30 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)
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 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).