Research map / Label-efficient and robust learning / Recognition and long-tail
ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ProtoArgNet, which combines prototypical parts into class-specific super-prototypes via an MLP-like layer over channel-wise-max activation maps, then sparsifies it into a QBAF for supporting and attacking explanations.
Problem
Prototypical-part-learning image classifiers associate each class with multiple prototypes and cannot capture spatial relationships among prototypical parts, while their explanations are complex.
Approach
Class-specific super-prototypes combine localized prototype activations (channel-wise max) through an MLP-like layer over activation maps, and the resulting network is sparsified and translated into a quantitative bipolar argumentation framework (QBAF) to provide 'looks like that' and 'differs from that' explanations.
Key results
- ProtoArgNet reaches 85.4% accuracy on CUB and 89.3% on Cars, outperforming ST-ProtoPNet (83.6%, 88.7%) and other prototypical-part baselines.
- On the synthetic SHAPES dataset, ProtoArgNet achieves 99.8% accuracy while all baselines stay near 50%.
- Average activated prototypes per example drop to 24.57 on CUB and 8.42 on Cars, versus over 1000 for ProtoPNet-style methods, with accuracy unchanged up to 80% sparsification.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ProtoArgNet instead of ST-ProtoPNet | image classification · CUB | accuracy: 83.6 ± 0.2 -> 85.4 ± 0.2 | Helped |
| use ProtoArgNet instead of ProtoPool | image classification · Cars | accuracy: 88.9 ± 0.1 -> 89.3 ± 0.3 | Helped |
| use ProtoArgNet instead of ST-ProtoPNet | image classification · Brain Tumor MRI | accuracy: 98.7 ± 0.3 -> 99.5 ± 0.3 | Helped |
| use ProtoArgNet instead of TesNet | image classification · SHAPES | accuracy: 50.6 ± 0.3 -> 99.8 ± 0.1 | Helped |
| add cosine similarity instead of L2 distance to the prototype layer (ablation) | ProtoArgNet with super-prototypes and SMLP classifier · CUB | accuracy: 82.7 -> 85.4 | Helped |
| add cosine similarity instead of L2 distance to the prototype layer (ablation) | ProtoArgNet with super-prototypes and SMLP classifier · Cars | accuracy: 88.3 -> 89.3 | Helped |
| add super-prototypes (ablation) | SMLP classifier with cosine prototype layer · CUB | accuracy: 81.9 -> 85.4 | Helped |
| add super-prototypes (ablation) | SMLP classifier with cosine prototype layer · Cars | accuracy: 88.0 -> 89.3 | Helped |
| add SMLP classifier instead of a fixed classifier (ablation) | ProtoArgNet with super-prototypes and cosine prototype layer · CUB | accuracy: 83.5 -> 85.4 | Helped |
| add SMLP classifier instead of a fixed classifier (ablation) | ProtoArgNet with super-prototypes and cosine prototype layer · Cars | accuracy: 88.9 -> 89.3 | Helped |
| set sparsification ratio to 0.8 instead of 0.4 (ablation) | ProtoArgNet · CUB | accuracy: 85.4 -> 85.4 | No clear effect |
| set sparsification ratio to 0.8 instead of 0.4 (ablation) | ProtoArgNet · Cars | accuracy: 89.3 -> 89.3 | No clear effect |
Surprising findings
- Even with a fixed logistic regression layer instead of the SMLP super-prototype layers, ProtoArgNet still beats prior state-of-the-art methods on CUB and Cars.
Limitations
- Accuracy drops at 94% sparsification ratio on CUB and Cars.
- Evaluation on a synthetic SHAPES dataset designed to test spatial relations, which may not reflect real-world difficulty.
- Explanations still require users to trace QBAF relations; human-readable interpretation is left for future work.
Experimental setup
Datasets: CUB-200-2011 Stanford Cars Brain Tumor MRI SHAPES (adapted binary, 10,000 images)
Benchmarks: CUB-200-2011 Stanford Cars Brain Tumor MRI SHAPES
Metrics: top-1 accuracy average number of activated prototypes (#AAP)
Baselines: ProtoPNet ProtoTrees ProtoPShare ProtoPool TesNet D-ProtoPNet ST-ProtoPNet PIP-Net
Builds on: ProtoPNet SpArX ProtoSpArX QBAF
Closest papers by meaning
- Interpretable Image Classification with Adaptive Prototype-based Vision Transformers (NeurIPS 2024 · similarity 0.89)
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.89)
- Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.88)
- Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · similarity 0.87)
- LucidPPN: Unambiguous Prototypical Parts Network for User-centric Interpretable Computer Vision (ICLR 2025 · similarity 0.86)
- MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.