Research map / Vision-language and multimodal / Human and embodied modeling
Visual Explanations of Image-Text Representations via Multi-Modal Information Bottleneck Attribution
Ying Wang, Tim G. J. Rudner, Andrew Gordon Wilson
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 |
|---|---|---|---|
| Use M2IB instead of the strongest baseline on MS-CXR image localization | M2IB · MS-CXR image | % localization IoU: 21.65 -> 22.59 | Improved |
| Use M2IB instead of Saliency on MS-CXR image ROAR+ | M2IB · MS-CXR image | % ROAR+: 25.46 -> 38.7 | Improved |
| Use M2IB instead of the transformer-specific method on Conceptual Captions text ROAR+ | M2IB · Conceptual Captions text | % ROAR+: 53.57 -> 60.41 | Improved |
| Use M2IB instead of the transformer-specific method on Conceptual Captions image confidence increase | M2IB · Conceptual Captions image | % confidence increase: 37.41 -> 41.55 | Improved |
| Use M2IB instead of Kernel SHAP on Conceptual Captions text confidence increase | M2IB · Conceptual Captions text | % confidence increase: 46.87 -> 38.55 | Hurt |
| Use M2IB instead of RISE on MS-CXR text confidence increase | M2IB · MS-CXR text | % confidence increase: 57.2 -> 35.48 | Hurt |
| Use M2IB instead of RISE on MS-CXR text confidence drop | M2IB · MS-CXR text | % confidence drop: 1.16 -> 2.28 | Hurt |
| Use M2IB instead of RISE on Conceptual Captions image confidence drop | M2IB · Conceptual Captions image | % confidence drop: 1.12 -> 1.11 | No clear effect |
| Use M2IB instead of Saliency on MS-CXR image confidence drop | M2IB · MS-CXR image | % confidence drop: 0.81 -> 0.55 | Improved |
| Choose layer 9, beta=0.1, sigma^2=1 instead of the worst grid setting (ablation) | M2IB · Conceptual Captions image | % ROAR+: 6.42 -> 12.72 | Improved |
| Choose the best versus worst hyperparameters (ablation) | M2IB · MS-CXR image | % ROAR+: 25.20 -> 37.17 | Improved |
| Choose layer 9 and moderate beta for text on Conceptual Captions (ablation) | M2IB · Conceptual Captions text | % ROAR+: 8.46 -> 18.16 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GradCAM — Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization (ICCV 2017) · better on 5 of 5 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
- Narrowing Information Bottleneck Theory for Multimodal Image-Text Representations Interpretability (ICLR 2025 · similarity 0.89)
- Attribution Analysis Meets Model Editing: Advancing Knowledge Correction in Vision Language Models with VisEdit (AAAI 2025 · similarity 0.87)
- Machine Vision Therapy: Multimodal Large Language Models Can Enhance Visual Robustness via Denoising In-Context Learning (ICML 2024 · similarity 0.87)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.87)
- VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety (ICLR 2026 · similarity 0.87)
- Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers (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.
