Research map / Architectures and efficiency / Attention and transformers
Transformer Interpretability Beyond Attention Visualization
Hila Chefer, Shir Gur, Lior Wolf
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 |
|---|---|---|---|
| Use class-specific relevance and gradient propagation integrated across attention blocks | ViT-base explanations · ImageNet-Segmentation | mIoU: 57.94 -> 61.95 | Improved |
| Use class-specific relevance and gradient propagation integrated across attention blocks | ViT-base explanations · ImageNet val, negative perturbation, predicted class | AUC: 53.1 -> 54.16 | Improved |
| Use class-specific relevance and gradient propagation integrated across attention blocks | ViT-base explanations · ImageNet val, positive perturbation, predicted class | AUC: 19.64 -> 17.03 | Improved |
| Use class-specific relevance and gradient propagation integrated across attention blocks | ViT-base explanations · ImageNet val, negative perturbation, target class | AUC: 50.49 -> 55.04 | Improved |
| Use class-specific relevance and gradient propagation integrated across attention blocks | ViT-base explanations · ImageNet val, positive perturbation, target class | AUC: 19.64 -> 16.04 | Improved |
| Replace attention-map gradients with attention maps (ablation) | Full method · ImageNet-Segmentation | segmentation performance | Hurt |
| Add rollout across attention blocks rather than use only the output-nearest block (ablation) | Full method · ImageNet-Segmentation and ImageNet predicted-class perturbation | performance | Improved |
| Use the output-nearest attention block rather than the input-nearest block (ablation) | Single-block relevance-gradient explanation · ImageNet visualizations | informativeness | Improved |
Closest papers by meaning
- ICEv2: Interpretability, Comprehensiveness, and Explainability in Vision Transformer (IJCV 2024 · similarity 0.88)
- Graph Convolutions Enrich the Self-Attention in Transformers! (NeurIPS 2024 · similarity 0.87)
- RMT: Retentive Networks Meet Vision Transformers (CVPR 2024 · similarity 0.87)
- AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers (ICML 2024 · similarity 0.86)
- Attention Guided CAM: Visual Explanations of Vision Transformer Guided by Self-Attention (AAAI 2024 · similarity 0.86)
- Rethinking Attention: Exploring Shallow Feed-Forward Neural Networks as an Alternative to Attention Layers in Transformers (Student Abstract) (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.
