Research map / Detection and segmentation / Open-set and few-shot detection
Efficient Unsupervised Shortcut Learning Detection and Mitigation in Transformers
Lukas Kuhn, Sari Sadiya, Jörg Schlötterer, Florian Buettner, Christin Seifert, Gemma Roig
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 ASM (patch ablation + last-layer retraining) to the fine-tuned ViT | ViT-B/16 baseline · ISIC | worst-group accuracy (%): 51.7 -> 74.7 | Improved |
| add ASM to the fine-tuned ViT | ViT-B/16 baseline · Knee Radiographs | worst-group accuracy (%): 37.3 -> 81.7 | Improved |
| add ASM to the fine-tuned ViT | ViT-B/16 baseline · ImageNet-W | worst-group accuracy (%): 69.0 -> 87.0 | Improved |
| add last-layer retraining after patch ablation (ablation) | ASM · ISIC | worst-group accuracy (%): 61.0 -> 74.7 | Improved |
| add last-layer retraining after patch ablation (ablation) | ASM · ImageNet-W | worst-group accuracy (%): 75.0 -> 87.0 | Improved |
| add last-layer retraining after patch ablation (ablation) | ASM · Knee Radiographs | worst-group accuracy (%): 80.0 -> 81.7 | No clear effect |
| patch ablation only, no retraining, versus baseline (ablation) | ASM w/o retraining · Knee Radiographs | worst-group accuracy (%): 37.3 -> 80.0 | Improved |
| use unsupervised ASM instead of group-annotated DFR | DFR · ISIC | worst-group accuracy (%): 59.3 -> 74.7 | Improved |
| use subsampling groups (SUBG) instead of baseline | ViT-B/16 baseline · ISIC | average-group accuracy (%): 86.2 -> 74.3 | Hurt |
| use 3 clusters instead of 2 (overclustering) (ablation) | ASM w/o retraining · ISIC | worst-group accuracy (%): 61.0 -> 56.7 | No clear effect |
| use ASM instead of JTT for runtime | JTT · ISIC | runtime (s): 2886.3 -> 121.1 | Improved |
Closest papers by meaning
- Transformers are almost optimal metalearners for linear classification (NeurIPS 2025 · similarity 0.85)
- In-context Convergence of Transformers (ICML 2024 · similarity 0.84)
- From Shortcut to Induction Head: How Data Diversity Shapes Algorithm Selection in Transformers (NeurIPS 2025 · similarity 0.84)
- Trained Transformer Classifiers Generalize and Exhibit Benign Overfitting In-Context (ICLR 2025 · similarity 0.83)
- Transformer Learns Optimal Variable Selection in Group-Sparse Classification (ICLR 2025 · similarity 0.83)
- Transformers Learn Low Sensitivity Functions: Investigations and Implications (ICLR 2025 · similarity 0.82)
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.
