Research map / Label-efficient and robust learning / Recognition and long-tail
Debugging Concept Bottleneck Models through Removal and Retraining
Eric Enouen, sainyam galhotra
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 |
|---|---|---|---|
| remove expert-marked spurious concepts without retraining (ablation) | Post-hoc CBM · Waterbirds (user feedback) | worst-group accuracy: 25.8 -> 13.9 | Hurt |
| remove expert-marked spurious concepts without retraining (ablation) | PIP-Net · Waterbirds (user feedback) | worst-group accuracy: 71.9 -> 74.4 | Improved |
| naive retraining after removal on the biased dataset (ablation) | PIP-Net · Waterbirds (user feedback) | worst-group accuracy: 71.9 -> 72.5 | No clear effect |
| add CBDebug (reweight and augment) retraining | PIP-Net · Waterbirds (user feedback) | worst-group accuracy: 71.9 -> 79.4 | Improved |
| add CBDebug retraining | Post-hoc CBM · Waterbirds (user feedback) | worst-group accuracy: 25.8 -> 51.9 | Improved |
| add CBDebug retraining | Post-hoc CBM · CelebA (automated feedback) | worst-group accuracy: 8.7 -> 51.3 | Improved |
| add augmentation on top of reweighting (ablation) | CBDebug on Post-hoc CBM · Waterbirds (user feedback) | worst-group accuracy: 55.6 -> 51.9 | Hurt |
| add reweighting on top of augmentation (ablation) | CBDebug on PIP-Net · Waterbirds (user feedback) | worst-group accuracy: 75.5 -> 79.4 | Improved |
| add reweighting on top of augmentation (ablation) | CBDebug on Post-hoc CBM · ISIC (automated feedback) | AUROC: 18.6 -> 58.0 | Improved |
| use reweighting only (no augmentation) (ablation) | CBDebug on Post-hoc CBM · MetaShift (automated feedback) | worst-group accuracy: 84.5 -> 84.1 | No clear effect |
| CBDebug removal replaces spurious concepts with robust ones | Post-hoc CBM · Waterbirds concept inspection | top concepts (qualitative) | Improved |
Closest papers by meaning
- Editable Concept Bottleneck Models (ICML 2025 · similarity 0.90)
- Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models (NeurIPS 2026 · similarity 0.89)
- Learning to Intervene on Concept Bottlenecks (ICML 2024 · similarity 0.89)
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · similarity 0.88)
- An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations (NeurIPS 2025 · similarity 0.88)
- VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · 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.
