Research map / Architectures and efficiency / Attention and transformers
LeGrad: An Explainability Method for Vision Transformers via Feature Formation Sensitivity
Walid Bousselham, Angie Boggust, Sofian Chaybouti, Hendrik Strobelt, Hilde Kuehne
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 |
|---|---|---|---|
| add ReLU on gradients (ablation) | LeGrad (L/14, last layer) · ImageNet-val perturbation negative | AUC: 47.81 -> 49.32 | Improved |
| add ReLU on gradients (ablation) | LeGrad (L/14, all layers) · ImageNet-val perturbation negative | AUC: 52.01 -> 54.48 | Improved |
| add ReLU on gradients (ablation) | LeGrad (L/14, all layers) · ImageNet-val perturbation positive | AUC (lower better): 16.80 -> 15.23 | Improved |
| use intermediate tokens from all layers (ablation) | LeGrad (L/14, with ReLU) · ImageNet-val perturbation negative | AUC: 49.32 -> 54.48 | Improved |
| use intermediate tokens from all layers (ablation) | LeGrad (L/14, with ReLU) · ImageNet-val perturbation positive | AUC (lower better): 19.95 -> 15.23 | Improved |
| use intermediate tokens from all layers (ablation) | LeGrad (H/14, with ReLU) · ImageNet-val perturbation positive | AUC (lower better): 19.50 -> 18.26 | Improved |
| include [CLS] token in gradient computation (ablation) | GradCAM (ViT-B/16) · OpenImagesV7 | mIoU: 8.18 -> 8.72 | Improved |
| add LeGrad-style multi-layer aggregation (ablation) | GradCAM (ViT-B/16) · OpenImagesV7 | mIoU: 8.72 -> 9.51 | Improved |
| add LeGrad-style multi-layer aggregation (ablation) | GradCAM (ViT-B/16) · ImageNet-val perturbation negative | AUC: 45.26 -> 45.31 | No clear effect |
| use LeGrad instead of TextSpan | OpenCLIP ViT-B/16 · ImageNet-val perturbation negative target | AUC: 52.81 -> 52.27 | Hurt |
| apply LeGrad to ImageBind | ImageBind · ADE20K Sound Prompted | mAP: 18.3 -> 48.0 | Improved |
| apply LeGrad to ImageBind | ImageBind · ADE20K Speech Prompted | mAP: 20.2 -> 23.3 | Improved |
Later papers that compared with this method
1 later accepted paper reports LeGrad as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- DAVE: Distribution-Aware Attribution via ViT Gradient Decomposition (ICML 2026 · better on 8 of 8 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Attention Guided CAM: Visual Explanations of Vision Transformer Guided by Self-Attention (AAAI 2024 · similarity 0.88)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.88)
- On the Faithfulness of Vision Transformer Explanations (CVPR 2024 · similarity 0.87)
- Token Transformation Matters: Towards Faithful Post-hoc Explanation for Vision Transformer (CVPR 2024 · similarity 0.86)
- Improving Interpretation Faithfulness for Vision Transformers (ICML 2024 · similarity 0.86)
- Beyond the Doors of Perception: Vision Transformers Represent Relations Between Objects (NeurIPS 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.
