Research map / Architectures and efficiency / Attention and transformers
Learning to Estimate Shapley Values with Vision Transformers
Ian Connick Covert, Chanwoo Kim, Su-In Lee
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 |
|---|---|---|---|
| Train ViT with random attention masking / fine-tune with masking (vs no fine-tuning) (ablation) | ViT classifier handling held-out patches · ImageNette, MURA | KL divergence to full-image prediction under patch removal | Improved |
| Use learned ViT explainer (ViT Shapley) instead of baselines | ViT explanations, target class · ImageNette and MURA insertion/deletion/faithfulness | insertion, deletion, faithfulness | Improved |
| Explain non-target classes with ViT Shapley vs baselines | ViT explanations, non-target classes · ImageNette | insertion and faithfulness | Improved |
| Add extra attention block and full fine-tuning to explainer (config D vs config A) (ablation) | ViT Shapley explainer architecture · ImageNette | validation loss: 4.309 vs 4.332 | Improved |
| Train explainer from scratch instead of fine-tuning ViT (ablation) | ViT Shapley explainer · ImageNette | validation loss: 4.331 vs 4.309 for best fine-tuned config | Hurt |
| Use attention-based explanations (attention last, rollout) for ViT | ViT explanations · ImageNette non-target classes | insertion/deletion/faithfulness | No clear effect |
| Gradient-based methods (GradCAM, SmoothGrad, VarGrad) vs random baseline | ViT explanations, target class · ImageNette insertion | insertion | No clear effect |
Closest papers by meaning
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.87)
- On the Faithfulness of Vision Transformer Explanations (CVPR 2024 · similarity 0.86)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.86)
- SHAP values via sparse Fourier representation (NeurIPS 2025 · similarity 0.85)
- Gnothi Seauton: Empowering Faithful Self-Interpretability in Black-Box Transformers (ICLR 2025 · similarity 0.85)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (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.
