Research map / Architectures and efficiency / CNN design
One Explanation is Not Enough: Structured Attention Graphs for Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Present structured attention graphs instead of a single saliency map | VGGNet explanations · User study, comparative counterfactual questions | correct responses: 5.4 -> 8.6 vs I-GOS; 5.3 -> 8.6 vs Grad-CAM | Helped |
| Use beam search to find compact sufficient explanations | VGGNet · ImageNet validation | images explained with at most 10 patches: about 50% -> about 80% | Helped |
| Remove click interaction from SAG (ablation) | SAG · User study, comparative counterfactual questions | correct responses | Hurt |
| Remove graph visualization, retaining root nodes and interaction (ablation) | SAG · User study, comparative counterfactual questions | correct responses | Hurt |
| Use SAG without click interaction rather than Grad-CAM or I-GOS (ablation) | SAG/I · User study, comparative counterfactual questions | correct responses | Helped |
| Use blurred pixels rather than zeros for perturbed image regions (ablation) | MSE search · ImageNet validation | coverage of images explained at a given MSE size | Helped |
| Use ResNet rather than VGGNet as the classifier being explained | BeamS-3 · ImageNet validation, overlap = 1 | mean number of diverse MSEs: 4.18 -> 3.27 | Hurt |
| Use ResNet rather than VGGNet as the classifier being explained | BeamS-15 · ImageNet validation, overlap = 1 | mean number of diverse MSEs: 4.51 -> 3.41 | Hurt |
Closest papers by meaning
- Attention Branch Network: Learning of Attention Mechanism for Visual Explanation (CVPR 2019 · similarity 0.88)
- Learn to Pay Attention (ICLR 2018 · similarity 0.88)
- Explainability Methods for Graph Convolutional Neural Networks (CVPR 2019 · similarity 0.87)
- Explicitly Modeled Attention Maps for Image Classification (AAAI 2021 · similarity 0.87)
- Graph Attention is Not Always Beneficial: A Theoretical Analysis of Graph Attention Mechanisms via Contextual Stochastic Block Models (ICML 2025 · similarity 0.86)
- Structure Your Data: Towards Semantic Graph Counterfactuals (ICML 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.