Research map / Label-efficient and robust learning / Recognition and long-tail
Editable Concept Bottleneck Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Editable Concept Bottleneck Models (ECBMs), which use influence-function closed-form approximations (with EK-FAC acceleration) to edit trained CBMs at concept-label, concept, and data levels without retraining.
Problem
Trained concept bottleneck models cannot efficiently remove or correct training data, concept labels, or concepts (for privacy, mislabelling, spurious concepts, annotation errors) without costly retraining from scratch.
Approach
Applies influence functions to the two-stage CBM (concept predictor then linear label predictor) to derive closed-form Newton-style edit updates for three editing levels, handles concept-dimension changes by zero-padding and mapping, and accelerates Hessian inversion with EK-FAC.
Key results
- On OAI concept-label editing, ECBM reaches F1 0.8808 vs 0.8825 for retraining while runtime drops from 297.77 to 2.36 minutes.
- On CUB, ECBM runtime is 0.65 minutes vs 85.56 for retraining with F1 0.7963 vs 0.7971.
- Mean RMIA score of removed members drops from 0.063505 to 0.052105 after ECBM editing, close to non-members' 0.049465.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ECBM instead of retraining | CBM · Concept-label edit, OAI | F1 score: 0.8825 -> 0.8808 | No clear effect |
| Use ECBM instead of retraining | CBM · Concept-label edit, CUB | F1 score: 0.7971 to 0.7963 | No clear effect |
| Use ECBM instead of retraining | CBM · Concept-label edit, CelebA | F1 score: 0.3827 -> 0.3845 | Mixed |
| Accelerate the editing method with EK-FAC | CBM-IF · Concept-label edit, OAI | RT (minute): 4.63 -> 2.36 | Helped |
| Accelerate the editing method with EK-FAC | CBM-IF · Concept-label edit, CUB | F1 score: 0.7699$\pm$0.0035 -> 0.7963$\pm$0.0050 | Helped |
| Use ECBM instead of retraining | CBM · Concept edit, OAI | F1 score: 0.8448 -> 0.8403 | No clear effect |
| Use ECBM instead of retraining | CBM · Data edit, CUB | F1 score: 0.7838 -> 0.7827 | No clear effect |
| Use ECBM instead of retraining | CBM · Data edit, CelebA | F1 score: 0.3797 -> 0.3748 | Hurt |
| Use EK-FAC-accelerated ECBM instead of retraining | CBM · Concept-label edit, OAI | RT (minute): 297.77 -> 2.36 | Helped |
| Use EK-FAC-accelerated ECBM instead of retraining | CBM · Concept-label edit, CUB | RT (minute): 85.56 -> 0.65 | Helped |
| Remove most influential concepts rather than least influential concepts | ECBM on CUB · Concept edit, CUB | F1 score difference: more than 0.025; below 0.005 | Mixed |
| Remove 200 training samples with ECBM | CBM on CUB · Data edit, CUB | RMIA score: 0.063505 -> 0.052105 | Helped |
Surprising findings
- The original Hessian-based CBM-IF is notably less accurate than the EK-FAC-accelerated ECBM, which the authors attribute to ill-conditioned or non-convex Hessians.
Limitations
- Hessian may be non-convex, ill-conditioned or singular, so inverse-based updates can be imprecise; EK-FAC is used as an approximation.
- Concept-level editing is restricted to concept removal; insertion is only claimed as an extension.
- Evaluation uses only a few datasets and the label predictor is assumed linear.
Experimental setup
Datasets: OAI CUB CelebA
Benchmarks: OAI CUB CelebA
Metrics: F1 score runtime (minutes) RMIA score
Baselines: Retrain CBM-IF
Builds on: concept bottleneck models influence functions EK-FAC label-free CBM
Closest papers by meaning
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · similarity 0.90)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.90)
- Bayesian Concept Bottleneck Models with LLM Priors (NeurIPS 2025 · similarity 0.89)
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.89)
- A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · similarity 0.89)
- DCBM: Data-Efficient Visual Concept Bottleneck Models (ICML 2025 · similarity 0.88)
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.