Research map / Label-efficient and robust learning / Recognition and long-tail
ProtoShotXAI: Using Prototypical Few-Shot Architecture for Explainable AI
Samuel Hess, Gregory Ditzler
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| average the support set into a class prototype instead of taking the best-matching single exemplar (ablation) | ProtoShotXAI · MNIST revolving six | agreement with model predicted class | Improved |
| multiply features by classification weights of the class (ablation) | ProtoShotXAI · feature attribution maps | negative attributions | Improved |
| use Euclidean distance instead of cosine similarity (ablation) | ProtoShotXAI · revolving six | interpretability and consistency with model response | Hurt |
| compute distance at the final classification layer instead of the class-weighted feature layer (ablation) | ProtoShotXAI · feature attribution maps | attribution interpretability | Hurt |
| add few-shot training of adaptive weights between prototype and query features (ablation) | ProtoShotXAI · feature attribution maps | attribution quality | No clear effect |
| increase support set size from 1 to about 60 samples (ablation) | ProtoShotXAI · ImageNet attribution (elephant and zebra) | attribution stability: 1 to 220 samples varied; negligible difference by 60; 100 used | Improved |
| use zero reference pixel instead of average value reference (ablation) | ProtoShotXAI · ImageNet tank/bulldog image | attribution magnitude in background | Mixed |
| increase perturbation pad size from 1 to 50 (ablation) | ProtoShotXAI · ImageNet attribution | attribution coverage vs detail: pad 1 to 50; pad of 15 used | Mixed |
| use ResNet50 with Stylized-ImageNet weights instead of ImageNet weights | ResNet50 · tank/bulldog image | evidence of shape bias in attributions: 74.9% -> 43.4% accuracy | No clear effect |
| threshold ProtoShotXAI in-class score to detect adversarial samples | ProtoShotXAI · Adversarial MNIST (FGSM, 1,000 samples) | detection ROC | Improved |
Closest papers by meaning
- Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics (NeurIPS 2024 · similarity 0.88)
- Gnothi Seauton: Empowering Faithful Self-Interpretability in Black-Box Transformers (ICLR 2025 · similarity 0.87)
- Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · similarity 0.86)
- Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition (ICLR 2024 · similarity 0.85)
- F-Fidelity: A Robust Framework for Faithfulness Evaluation of Explainable AI (ICLR 2025 · similarity 0.85)
- Pantypes: Diverse Representatives for Self-Explainable Models (AAAI 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.
