Research map / Label-efficient and robust learning / Recognition and long-tail
Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning
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.
Shows that ProtoPNet's pixel-level attribution maps violate faithfulness axioms and proposes ProtoPFaith, which computes Shapley-value explanations of prototype similarities via DASP on a probabilistic ProtoPNet.
Problem
Pixel-level explanations of ProtoPNet-style case-based reasoning models are assumed faithful, but the attribution maps used to explain prototype similarities may violate established explanation axioms.
Approach
Drop ProtoPNet's log-activation to restore linearity, convert the trained ProtoPNet into a lightweight probabilistic model with closed-form moments for squared L2-distance and ReLU1, and approximate Shapley values of each prototype's similarity with DASP to yield faithful pixel attributions.
Key results
- ProtoPFaith outperforms ProtoPNet on AOPC by a factor greater than 10^3 across all experiments (e.g., CUB-200-2011 ResNeXt50: -3.72 vs -0.00045).
- Test balanced accuracy of ResNeXt50 is 72.0% on CUB-200-2011 and 83.4% on Stanford Dogs; ConvNet reaches 72.6% on RSNA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProtoPFaith Shapley-value explanations instead of ProtoPNet attribution maps | ProtoPNet with ResNeXt50 · CUB-200-2011 | AOPC: -0.000446 -> -3.721184 | Helped |
| Use ProtoPFaith Shapley-value explanations instead of ProtoPNet attribution maps | ProtoPNet with ResNeXt50 · Stanford Dogs | AOPC: -0.001767 -> -1.850909 | Helped |
| Use ProtoPFaith Shapley-value explanations instead of ProtoPNet attribution maps | ProtoPNet with ConvNet · RSNA | AOPC: -0.001360 -> -8.816450 | Helped |
Surprising findings
- ProtoPNet's attribution maps are proven to violate Completeness, Dummy, Sensitivity, and Linearity axioms despite case-based reasoning being faithful.
- ResNet on RSNA showed collapsed prototypes per class yet still achieved 79.8% test BAcc.
Limitations
- Currently only applied to single-label image classification.
- Shapley-based explanations are harder to interpret and less sparse than ProtoPNet's crops.
- Explanations include large background portions, making them less human-understandable.
Experimental setup
Datasets: CUB-200-2011 Stanford Dogs RSNA pneumonia detection (subset)
Benchmarks: CUB-200-2011 Stanford Dogs RSNA
Metrics: balanced accuracy Area Over the Perturbation Curve (AOPC)
Baselines: ProtoPNet
Builds on: ProtoPNet DASP Shapley values AOPC
Closest papers by meaning
- ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation (AAAI 2025 · similarity 0.87)
- MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · similarity 0.86)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.86)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.86)
- Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.86)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.86)
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.