Research map / Architectures and efficiency / Attention and transformers
CAR-MIL: Counterfactual Attention Regularization for Multiple Instance Learning
Imane Chraki, Pierre Marza, Stergios Christodoulidis, Maria Vakalopoulou
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 counterfactual attention regularization (L1) | ABMIL · MNIST-bags Adjacent Pairs | instance AUPRC+/-: 61.5 ± 0.9 -> 84.6 ± 5.2 | Improved |
| add counterfactual attention regularization (L1) | ABMIL · MNIST-bags Four Bags | instance AUPRC+/-: 53.1 ± 0.1 -> 89.7 ± 0.9 | Improved |
| add counterfactual attention regularization (cosine) | ABMIL · TCGA-LUAD (TP53) | AUC: 74.0±4.3 -> 78.1±4.1 | Improved |
| add counterfactual attention regularization (cosine) | ABMIL · BRACS | balanced accuracy: 38.2±4.1 -> 43.1±2.0 | Improved |
| add counterfactual attention regularization | ABMIL · TCGA-BRCA | AUC: 95.3±1.7 -> 95.5±2.2 | No clear effect |
| add counterfactual attention regularization to DSMIL (ablation) | DSMIL · TCGA-LUAD (TP53) | AUC: 66.9±4.7 -> 71.2±4.1 | Improved |
| add counterfactual attention regularization to TransMIL (ablation) | TransMIL · TCGA-LUAD (TP53) | AUC: 71.7±5.4 -> 73.6±4.4 | Improved |
| add counterfactual attention regularization to TransMIL (ablation) | TransMIL · BRACS | balanced accuracy: 40.0±3.6 -> 42.6±6.1 | Improved |
| add counterfactual attention regularization to TransMIL (ablation) | TransMIL · CAMELYON16 | AUPRC+: 28.2±3.6 -> 32.6±0.3 | Improved |
| add counterfactual attention regularization to DSMIL (ablation) | DSMIL · CAMELYON16 | AUPRC+: 80.6±11.2 -> 87.8±2.9 | Improved |
| use cosine instead of L1 attention-logit distance (ablation) | CAR-MIL · LUAD, BRACS, BRCA, NSCLC, CAMELYON16 | AUC/BACC: LUAD AUC 76.4±4.9 -> 78.1±4.1; CAMELYON16 AUPRC+ 94.8±1.0 -> 91.7±0.6 | Mixed |
| add the evidence-differential and proximity losses (alpha, lambda in 0.2, 0.8, 1.0) (ablation) | ABMIL · TCGA-BRCA | AUC: 94.5 -> 95.3 to 96.0 | Improved |
Closest papers by meaning
- Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging (CVPR 2025 · similarity 0.85)
- xMIL: Insightful Explanations for Multiple Instance Learning in Histopathology (NeurIPS 2024 · similarity 0.85)
- Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology (NeurIPS 2025 · similarity 0.85)
- Symb-xMIL: Symbolic Explanations for Multiple Instance Learning in Digital Pathology (NeurIPS 2026 · similarity 0.84)
- Do Multiple Instance Learning Models Transfer? (ICML 2025 · similarity 0.84)
- SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.84)
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.
