Research map / Label-efficient and robust learning / Recognition and long-tail
A Robust Prototype-Based Network with Interpretable RBF Classifier Foundations
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.
Analyzes deep prototype-based networks, shows CBC's probabilistic model has flaws, proposes a revised CBC head with trainable class-wise component priors and negative reasoning, and derives provable robustness bounds and a robustness loss.
Problem
Deep prototype-based networks (PBNs) are claimed to be interpretable, but their classification heads behave like RBF networks with unconstrained weights, so explanations can be misleading; the original Classification-by-Components (CBC) model also has scaling ambiguity and poor local minima.
Approach
Remove the importance variable from the original CBC probability tree and replace it with trainable class-wise component priors P(k|c), yielding an RBF-like head with sound negative reasoning; derive a robustness lower bound and use it as a loss to optimize adversarial robustness.
Key results
- Deep CBC with ConvNeXt-tiny reaches 87.8% on CUB, 93.0% on CARS and 93.9% on PETS, beating PIPNet (84.3/88.2/92.0).
- On MNIST shallow models, Robust CBC gets 15.2% certified robust accuracy at perturbation 1.0 versus 0.2% for non-robust CBC, while GLVQ obtains 32.3%.
- Shallow CBC TD reaches 95.9% test accuracy on MNIST, though its certified robustness is 0.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the PIPNet classification layer with the CBC head, enabling negative reasoning | PIPNet with ConvNeXt-tiny · CUB test | accuracy: 84.3±0.2 -> 87.8±0.1 | Helped |
| replace the PIPNet classification layer with the CBC head, enabling negative reasoning | PIPNet with ConvNeXt-tiny · CARS test | accuracy: 88.2±0.5 -> 93.0±0.0 | Helped |
| replace the PIPNet classification layer with the CBC head, enabling negative reasoning | PIPNet with ConvNeXt-tiny · PETS test | accuracy: 92.0±0.3 -> 93.9±0.1 | Helped |
| restrict CBC to positive reasoning only (ablation) | deep CBC · CUB test | accuracy: 87.8±0.1 -> 28.6±0.8 | Hurt |
| restrict CBC to positive reasoning only (ablation) | deep CBC · CARS test | accuracy: 93.0±0.0 -> 25.3±2.3 | Hurt |
| restrict CBC to positive reasoning only (ablation) | deep CBC · PETS test | accuracy: 93.9±0.1 -> 69.5±5.1 | Hurt |
| replace original CBC with the proposed CBC extension | shallow CBC · MNIST test | accuracy: 81.8±2.0 -> 87.4±0.3 | Helped |
| add negative reasoning rather than restricting reasoning to positive-only normalized weights (ablation) | shallow CBC · MNIST test | accuracy: 77.3±0.2 -> 87.4±0.3 | Helped |
| use Tangent Distance (ablation) | shallow CBC · MNIST test | accuracy: 87.4±0.3 -> 95.9±0.1 | Helped |
| optimize with the robustness loss (ablation) | shallow CBC · MNIST test | certified robust accuracy: 0.2±0.1 -> 15.2±1.7 | Helped |
| optimize with the robustness loss (ablation) | shallow CBC · MNIST test | accuracy: 87.4±0.3 -> 87.8±0.3 | No clear effect |
| combine the robustness loss with Tangent Distance (ablation) | shallow CBC · MNIST test | certified robust accuracy: 15.2±1.7 -> 1.6±0.2 | Hurt |
Surprising findings
- Original CBC allows infinitesimal reasoning evidence to yield confident predictions because of a scaling ambiguity in its probability tree.
- Empirical robustness of shallow models is good even without robustification, which the authors leave unexplained.
Limitations
- Deep model is only partially interpretable; component interpretation requires expert knowledge.
- Training is sensitive to pre-training, initialization and temperature choice; components can collapse to the same point.
- Certified robustness of tangent-distance variants is much weaker due to extra triangle-inequality step; experiments limited to image datasets.
Experimental setup
Datasets: CUB CARS PETS MNIST
Benchmarks: CUB CARS PETS MNIST
Metrics: test accuracy empirical robust accuracy certified robust accuracy Jensen-Shannon divergence
Baselines: PIPNet ProtoPool ProtoViT CBC pos. reas. GLVQ RBF network original CBC RBF-norm
Builds on: Classification-by-Components PIPNet ProtoPNet generalized learning vector quantization
Closest papers by meaning
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.88)
- MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · similarity 0.85)
- Post-hoc Part-Prototype Networks (ICML 2024 · similarity 0.85)
- Interpretable Image Classification with Adaptive Prototype-based Vision Transformers (NeurIPS 2024 · similarity 0.84)
- ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation (AAAI 2025 · similarity 0.84)
- Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.83)
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.