Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining
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 R3, a post-processing framework that uses a learned Bradley-Terry reward model over human ratings of prototype activations to reweigh, reselect, and retrain a pretrained ProtoPNet, improving both interpretability and accuracy.
Problem
ProtoPNet prototypes often rely on spurious background features and inconsistent parts, and their quality is not quantified by human preference, which hurts interpretability and sometimes accuracy.
Approach
Train a reward model from a small set of human 1-5 ratings of prototype activation patterns (converted to pairwise Bradley-Terry comparisons), then apply a reward-weighted prototype update, a reward-thresholded prototype reselection, and a standard ProtoPNet retraining step, without RL.
Key results
- R3-ProtoPNet improves CUB test accuracy over ProtoPNet across five backbones (e.g., VGG-19 with 10 prototypes per class: 77.58 to 79.60; DenseNet-121: 79.64 to 80.42) and the ensemble from 82.92 to 84.37.
- Average reward rises from 0.47 (ProtoPNet) to 0.60 (R3-ProtoPNet) and average activation precision from 73.22 to 86.83 on CUB.
- The reward model reaches 90.09% pairwise test accuracy using 500 human ratings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply reward reweighing and prototype reselection without retraining (ablation) | ProtoPNet · CUB-200-2011 test, ensemble | accuracy: 82.92 ± 0.09 -> 70.46 ± 0.82 | Hurt |
| add retraining after reward reweighing and prototype reselection (ablation) | R2-ProtoPNet · CUB-200-2011 test, ensemble | accuracy: 70.46 ± 0.82 -> 84.37 ± 0.20 | Helped |
| add reward reweighing and prototype reselection (ablation) | ProtoPNet · CUB-200-2011 test, average across architectures | reward: 0.47 -> 0.58 | Helped |
| reselect prototypes using the reward threshold (ablation) | ProtoPNet · CUB-200-2011 test, average across architectures | reward: 0.47 -> 0.53 | Helped |
| reweigh and reselect prototypes, then retrain | ProtoPNet · CUB-200-2011 test, average across architectures | activation precision: 73.22 -> 86.83 | Helped |
| apply R3 prototype updates and retraining | ProtoPNet · CUB-200-2011 test, DenseNet-161 | accuracy: 79.75 ± 0.27 -> 79.48 ± 0.36 | Hurt |
| reweigh and reselect prototypes, then retrain | ProtoPNet · CUB-200-2011 test, average across architectures | reward: 0.47 -> 0.60 | Helped |
| apply reward reweighing and prototype reselection without retraining (ablation) | ProtoPNet · Stanford Cars test, ensemble | accuracy: 90.39 ± 0.14 -> 73.42 ± 0.81 | Hurt |
| add retraining after reward reweighing and prototype reselection (ablation) | R2-ProtoPNet · Stanford Cars test, ensemble | accuracy: 73.42 ± 0.81 -> 91.57 ± 0.24 | Helped |
| reweigh and reselect prototypes, then retrain | ProtoPNet · Stanford Cars test, average across architectures | reward: 0.72 -> 0.86 | Helped |
| apply R3 prototype updates and retraining | ProtoPFormer · CUB-200-2011 test | accuracy: 84.27 ± 0.20 -> 85.31 ± 0.23 | Helped |
| add reward reweighing, prototype reselection, and retraining | ProtoPFormer · CUB-200-2011 test, global prototypes | activation precision: 83.48 -> 87.43 | Mixed |
Surprising findings
- The R2 update (reweighing plus reselection) alone sharply lowers test accuracy (e.g., VGG-19 with 10 prototypes drops to 50.41%), but retraining recovers and surpasses the original accuracy.
- DenseNet-161 is the one backbone where R3 does not improve test accuracy (79.75 for ProtoPNet vs 79.48 for R3-ProtoPNet).
Limitations
- Reward model trained only on individual image-prototype ratings, ignoring cross-image consistency; duplicate prototypes remain.
- Human ratings are subjective and noisy; the reward may penalize features useful for prediction but non-obvious to humans.
- Thresholds (reweigh, reselect, accept) are tuned per architecture and required manual tuning.
Experimental setup
Datasets: CUB-200-2011 Stanford Cars
Benchmarks: CUB-200-2011 Stanford Cars
Metrics: test accuracy average reward (learned reward model) activation precision
Baselines: ProtoPNet R2-ProtoPNet ProtoPFormer ProtoPNet with SDFA and SA
Builds on: ProtoPNet Bradley-Terry model ProtoPDebug reward-weighted regression
Closest papers by meaning
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.90)
- ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation (AAAI 2025 · similarity 0.88)
- LucidPPN: Unambiguous Prototypical Parts Network for User-centric Interpretable Computer Vision (ICLR 2025 · similarity 0.87)
- Rashomon Sets for Prototypical-Part Networks: Editing Interpretable Models in Real-Time (CVPR 2025 · similarity 0.87)
- Understanding Model Reprogramming for CLIP via Decoupling Visual Prompts (ICML 2025 · similarity 0.86)
- Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · 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.