Research map / Architectures and efficiency / Attention and transformers
From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models
Yearim Kim, Sangyu Han, Nojun Kwak
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 class-centering plus positive clamping to sharing ratios (vs Raw) (ablation) | SRD-input (VGG16) · ImageNet-S50 | Attribution Localization: .490 -> .561 | Improved |
| add class-centering plus positive clamping to sharing ratios (vs Raw) (ablation) | SRD-input (VGG16) · ImageNet-S50 | Sparseness: .746 -> .788 | Improved |
| add class-centering plus positive clamping to sharing ratios (vs Raw) (ablation) | SRD-input (ResNet50) · ImageNet-S50 | Sparseness: .596 -> .724 | Improved |
| add positive clamping after class-centering (vs Mean) (ablation) | SRD-input (VGG16) · ImageNet-S50 | Attribution Localization: .404 -> .561 | Improved |
| add positive clamping after class-centering (vs Mean) (ablation) | SRD-input (ResNet50) · ImageNet-S50 | Pointing Game: .958 -> .953 | Hurt |
| use SRD-high instead of LayerCAM-high | attribution method (VGG16) · ImageNet-S50 | Attribution Localization: .435 -> .566 | Improved |
| use SRD-input instead of GradCAM | attribution method (ResNet50) · ImageNet-S50 | Stability: .757 -> .104 | Improved |
| use SRD instead of Fullgrad on ReLU activation | ResNet50 CIFAR-100 · CIFAR-100 | Fidelity: 0.038 -> 0.078 | Improved |
| use SRD with Tanh activation instead of Fullgrad | ResNet50 CIFAR-100 · CIFAR-100 | Fidelity: 0.107 -> 0.115 | Improved |
| use input-scale SRD instead of high-layer SRD | SRD (VGG16) · ImageNet-S50 | Fidelity: .206 -> .069 | Hurt |
| use bisecting k-means instead of SAE for concept extraction at Layer1.1 | CAFE concept extraction · ResNet50 PFVs | Rel-l2 reconstruction error: 1.8952 -> 0.3651 | Improved |
| use SAE instead of bisecting k-means for concept extraction at Layer4.2 | CAFE concept extraction · ResNet50 PFVs | Rel-l2 reconstruction error: 0.5095 -> 0.4816 | Improved |
Closest papers by meaning
- Unsupervised Semantic Segmentation Facilitates Model Understanding (ECCV 2026 · similarity 0.89)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.89)
- Language Models Can Explain Visual Features via Steering (CVPR 2026 · similarity 0.89)
- Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · similarity 0.88)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.87)
- What's in the Image? A Deep-Dive into the Vision of Vision Language Models (CVPR 2025 · similarity 0.87)
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.
