Research map / Label-efficient and robust learning / Noisy labels
An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations
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.
Systematically measures how annotation noise in concept and target labels degrades concept bottleneck models (CBMs), identifies a susceptible concept subset, and mitigates via SAM training plus entropy-guided concept intervention at inference.
Problem
Concept bottleneck models rely on concept annotations that are often noisy, yet the impact of such noise on prediction performance, interpretability, and intervention effectiveness is not well understood.
Approach
Shows noise hurts CBMs disproportionately through a small "susceptible" concept set; mitigates with sharpness-aware minimization during training to stabilize these concepts, and at inference ranks concepts by predictive entropy to correct the most uncertain ones, using uncertainty as a proxy for susceptibility.
Key results
- On CUB, 40% concept noise drops task accuracy from 74.31% to 3.99% and CAS from 84.31% to 58.44%.
- SAM training improves task accuracy under 40% noise on AwA2 by 4.68% over the SGD baseline with a 0.60% concept accuracy gain.
- Combining SAM training with uncertainty-guided intervention reaches 85.1% task accuracy on CUB at 20% noise after 5 interventions (vs 54.2% without intervention).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase concept-label noise from 0% to 40% | Base CBM · CUB | task accuracy: 74.3% to 4.0% | Hurt |
| Increase concept-label noise from 0% to 40% | Base CBM · CUB | CAS: 84.3% to 58.4% | Hurt |
| Use concept-only noise instead of clean concept labels | CBM · CUB and AwA2 | task accuracy: 10% noise in CUB: 16.6% drop; 30% noise in CUB: 51% decline; 30% noise in AwA2: 9.4% drop | Hurt |
| Add SAM training instead of Base training | InceptionV3 CBM · CUB, 40% concept noise | task accuracy: 3.99 -> 4.95 | Helped |
| Add SAM training instead of Base training | InceptionV3 CBM · AwA2, 40% concept noise | task accuracy: 41.85 -> 46.53 | Helped |
| Add SAM training instead of Base training | InceptionV3 CBM · CUB, 40% concept noise | concept accuracy: +0.89 | Helped |
| Add SAM training instead of Base training | InceptionV3 CBM · CUB, 40% concept noise | CAS: +2.00 | Helped |
| Add SAM training | Susceptible concepts in the CBM · CUB, 20% concept noise | concept accuracy: +3.85% | Helped |
| Add SAM training | Susceptible concepts in the CBM · CUB, 40% concept noise | concept accuracy: +4.07% | Helped |
| Select concepts by uncertainty rather than at random | Base CBM · CUB, 20% concept noise, 5 interventions | task accuracy: 56.2 -> 71.2 | Helped |
| Select concepts by uncertainty rather than at random | SAM-trained CBM · CUB, 20% concept noise, 5 interventions | task accuracy: 59.9 -> 75.2 | Helped |
| Select concepts by uncertainty rather than at random | Base CBM · CUB, 40% concept noise, 5 interventions | task accuracy: 4.8 -> 6.9 | Helped |
Surprising findings
- Concept noise alone causes almost the same degradation as combined noise, while target noise has minimal effect, attributed to the linear target predictor.
- SAM gains vary across CBM variants and are marginal for AR-CBM and CEM (e.g., CEM at 40% noise 54.6 vs 54.5).
Limitations
- Assumes well-calibrated uncertainty estimates, which may fail in data-scarce or imbalanced regimes.
- Relies on a linear target predictor, which may limit generalization to complex decision boundaries.
- Restricted to binary concept labels.
Experimental setup
Datasets: CUB AwA2
Benchmarks: CUB AwA2
Metrics: task accuracy concept accuracy concept alignment score (CAS) intervention accuracy
Baselines: Base (SGD) Random intervention CCTP intervention label smoothing regularization symmetric cross-entropy CEM ECBM AR-CBM SCBM
Builds on: concept bottleneck models sharpness-aware minimization concept intervention
Closest papers by meaning
- Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.92)
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.90)
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · similarity 0.90)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.88)
- Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models (NeurIPS 2026 · similarity 0.88)
- A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · 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.