Research map / Architectures and efficiency / CNN design
Interpretability Benchmark for Evaluating Spatial Misalignment of Prototypical Parts Explanations
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.
Introduces a benchmark (PLC, PRC, PAC, AC metrics) that uses PGD adversarial modifications outside the top prototype's activation region to quantify spatial misalignment in prototypical-parts networks, and a spatial-alignment loss plus masking augmentation that reduces misalignment on ProtoPNet, TesNet, ProtoPool, and ProtoTree.
Problem
Prototypical-parts networks compute similarity maps at the penultimate layer, so the receptive field of a prototype's activation region can depend on image content outside that region, producing misleading explanations.
Approach
Adversarially perturb only the pixels outside a prototype's activation bounding box with PGD to test whether the activation changes, define location/activation/rank change metrics on that test, and train with a spatial alignment loss comparing similarity maps of the original image and a masked image (only the highly activated region kept), plus random masking augmentation.
Key results
- On CUB with ResNet-34, ProtoPNet's PAC drops from 23.7 (baseline) to 2.9 with lambda_align=10 and no masking augmentation, and PLC from 24.0 to 18.3 (table values for lambda=10, no MA).
- With lambda_align=10 plus masking augmentation, ProtoPNet reaches PLC 8.5, PRC 1.6, PAC 1.7 versus baseline 24.0, 13.5, 23.7 on CUB.
- Accuracy after adversarial modification for ProtoPNet with the best variant is 74.0% vs 68.2% baseline; the paper notes ProtoPool gains least and ProtoPNet gains most.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add alignment loss (λ=10) and masking augmentation | ProtoPNet · CUB-200-2011 test | PAC: 23.7 $\pm$ 2.8 -> 1.7 $\pm$ 0.7 | Helped |
| Add alignment loss (λ=10) and masking augmentation | ProtoPool · CUB-200-2011 test | PLC: 31.8 $\pm$ 0.8 -> 30.0 $\pm$ 1.6 | Helped |
| Add alignment loss (λ=1) and masking augmentation | ProtoTree · CUB-200-2011 test | PAC: 6.5 $\pm$ 0.1 -> 1.7 $\pm$ 0.5 | Helped |
| Add masking augmentation without alignment loss | TesNet · CUB-200-2011 test | PAC: 3.4 $\pm$ 0.3 -> 0.1 $\pm$ 0.0 | Helped |
| Add masking augmentation to alignment loss (λ=10) (ablation) | ProtoPNet · CUB-200-2011 test | PAC: 2.9 $\pm$ 1.1 -> 1.7 $\pm$ 0.7 | Helped |
| Add alignment loss (λ=10) to masking augmentation (ablation) | ProtoPNet · CUB-200-2011 test | PAC: 5.0 $\pm$ 0.7 -> 1.7 $\pm$ 0.7 | Helped |
| Remove TesNet orthogonality loss (ablation) | TesNet · CUB-200-2011 test | PAC: 3.4 $\pm$ 0.3 -> 2.2 $\pm$ 0.1 | Helped |
| Remove TesNet orthogonality and subspace-separation losses (ablation) | TesNet · CUB-200-2011 test | PAC: 3.4 $\pm$ 0.3 -> 1.6 $\pm$ 0.1 | Helped |
| Replace the binarized alignment mask with a smooth mask (ablation) | ProtoTree · CUB-200-2011 test | Accuracy before modification: 82.5 $\pm$ 0.2 -> 71.2 $\pm$ 4.8 | Hurt |
Surprising findings
- TesNet is markedly more robust to the benchmark than ProtoPNet, and ablating its orthogonality and subspace-separation losses does not remove this robustness, suggesting the similarity function is responsible.
- Smooth (non-binarized) masks during spatial-aligning training were generally ineffective and, for ProtoTree, harmed accuracy.
Limitations
- Spatial-alignment training adds an extra forward pass, increasing training time by about 40%.
- Benchmark evaluated mainly on CUB-200-2011 with ResNet-34; generalization shown only on VGG16 and Stanford Cars.
Experimental setup
Datasets: CUB-200-2011 Stanford Cars ImageNet
Benchmarks: CUB-200-2011 Stanford Cars
Metrics: PLC PRC PAC Accuracy Change
Baselines: ProtoPNet TesNet ProtoPool ProtoTree
Builds on: ProtoPNet TesNet ProtoPool ProtoTree projected gradient descent
Closest papers by meaning
- Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · similarity 0.86)
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.85)
- From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models (ICML 2026 · similarity 0.85)
- Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas (ICML 2025 · similarity 0.85)
- Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.85)
- Structure Your Data: Towards Semantic Graph Counterfactuals (ICML 2024 · similarity 0.84)
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.