Research map / Label-efficient and robust learning / Recognition and long-tail
Self-explaining and visual explanation networks
Builds networks that produce explanations through spatial attention, self-explaining design and explanation-based augmentation.
108 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 Benchmarks Tried Checklist Preprints Who Export
Where it stands
Quiet None of the tracked benchmarks has a report from 2025–26.
35 papers in 2023–24 and 31 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 2.8% to 2.4%.
No longer reported by papers here: Stanford Cars (accuracy, 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
- 2017 · Network Dissection: Quantifying Interpretability of Deep Visual Representations (CVPR 2017 · 1,404 citations)
- 2018 · Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV) (ICML 2018 · 477 citations)
- 2019 · This Looks Like That: Deep Learning for Interpretable Image Recognition (NeurIPS 2019 · 560 citations)
- 2020 · Interpretations are Useful: Penalizing Explanations to Align Neural Networks with Prior Knowledge (ICML 2020 · 98 citations)
- 2021 · Neural Prototype Trees for Interpretable Fine-Grained Image Recognition (CVPR 2021 · 257 citations)
- 2022 · A Framework for Learning Ante-Hoc Explainable Models via Concepts (CVPR 2022 · 43 citations)
- 2023 · Improving Interpretability of Deep Sequential Knowledge Tracing Models with Question-centric Cognitive Representations (AAAI 2023 · 88 citations)
- 2024 · Deep Unfolded Network with Intrinsic Supervision for Pan-Sharpening (AAAI 2024 · 37 citations)
- 2025 · Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation (CVPR 2025 · 14 citations)
- 2026 · From Attribution to Action: Jointly ALIGNing Predictions and Explanations (AAAI 2026 · 1 citation)
Earliest papers here
- Network Dissection: Quantifying Interpretability of Deep Visual Representations (CVPR 2017 · 1,404 citations)
- Visualizing Deep Neural Network Decisions: Prediction Difference Analysis (ICLR 2017 · 487 citations)
- Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV) (ICML 2018 · 477 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.
Stanford Cars · accuracy (higher is better)
4 papers, 2019–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 91.4 | This Looks Like That: Deep Learning for Interpretable Image Recognition (NeurIPS 2019 · 560 citations) |
| 2024 | 92.9 | Q-SENN: Quantized Self-Explaining Neural Networks (AAAI 2024 · 12 citations) |
What has been tried
86 papers here report 827 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 9%, hurt 22%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add the spatial consistency prior | QuickBird reduced-scale · ERGAS | 1.3766 -> 1.0865 | Helped | Deep Unfolded Network with Intrinsic Supervision for Pan-Sharpening (AAAI 2024 · 37 citations) |
| remove CKA loss and retain CCD loss | Caltech101 · accuracy | 93.88% -> 94.21% | No clear effect | MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · 16 citations) |
| Use the second predicted class as the comparison reference | Birds-525 · RD.@0.05 | 0.174 -> 0.198 | Helped | Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation (CVPR 2025 · 14 citations) |
| add a base template | ImageNet-1k, Wordnet (80k) concepts · F1-score | 0.3265 ± 0.009 -> 0.3361 ± 0.009 | Helped | WWW: A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts (CVPR 2024 · 13 citations) |
| Add quantization to Q-SENN | CUB-2011 · loc5 | 64.4 -> 78.1 | Helped | Q-SENN: Quantized Self-Explaining Neural Networks (AAAI 2024 · 12 citations) |
| Use a larger calibration set | MNIST and CIFAR-100 Super-class concept sets (ε=0.1) · prediction-set stability | Helped | Towards Modeling Uncertainties of Self-Explaining Neural Networks via Conformal Prediction (AAAI 2024 · 11 citations) | |
| add cosine similarity instead of L2 distance to the prototype layer | CUB · accuracy | 82.7 -> 85.4 | Helped | ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation (AAAI 2025 · 4 citations) |
| Replace CAPE explanation with μ-CAPE explanation | CUB200-2011 · mIoU | 13.4 -> 66.6 | Hurt | CAPE: CAM as a Probabilistic Ensemble for Enhanced DNN Interpretation (CVPR 2024 · 4 citations) |
| Add Saliency Partition (SP) to FG-VCE | CUB-200-2011, ResNet-50 · Ins. | 41.71 -> 43.02 | Helped | Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · 2 citations) |
| apply reward reweighing and prototype reselection without retraining | CUB-200-2011 test, ensemble · accuracy | 82.92 ± 0.09 -> 70.46 ± 0.82 | Hurt | Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · 2 citations) |
| Use only the primal loss, omitting stationarity and manifold loss terms | DSV generation · sample meaningfulness | Hurt | Deep Support Vectors (NeurIPS 2024 · 2 citations) | |
| Increase the number of concepts | CIFAR-100 · AUROC | Helped | Interpretable Failure Detection with Human-Level Concepts (AAAI 2025 · 1 citation) |
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
- ImageNet · 7 of 16 papers
- CUB-200 · 4 of 16 papers
- CIFAR-10 · 3 of 16 papers
- CUB-200-2011 · 3 of 16 papers
- CIFAR-100 · 2 of 16 papers
- ImageNet-1K · 2 of 16 papers
- Aircraft · 2 of 16 papers
- MNIST · 2 of 16 papers
Baselines
- GradCAM · 4 of 16 papers
- ScoreCAM · 4 of 16 papers
- Integrated Gradients · 3 of 16 papers
- LayerCAM · 2 of 16 papers
- GBP · 2 of 16 papers
- CRAFT · 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 66 papers; from OpenAlex and the papers' affiliations).
- Massachusetts Institute of Technology · 3 papers
- Korea University · 2 papers
- University of Cambridge · 2 papers
- Tsinghua University · 2 papers
- Chinese Academy of Sciences · 2 papers
- Stanford University · 2 papers
- Meta · 2 papers
- New Jersey Institute of Technology · 2 papers
- Harvard University · 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
- Prototype and concept dictionary models (104 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).