Research map / Label-efficient and robust learning / Recognition and long-tail
Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations
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 Energy-based Concept Bottleneck Models (ECBMs), which define joint energies over (input, concept, class) to unify concept prediction, concept intervention, and conditional concept interpretation as conditional probabilities, evaluated on CUB, CelebA, and AWA2.
Problem
Concept bottleneck models fail to capture high-order nonlinear interactions among concepts (so interventions do not propagate to correlated concepts) and cannot quantify complex conditional dependencies between concepts and class labels, while trailing black-box accuracy.
Approach
Use three energy networks (class x-y, concept x-c, global c-y) whose sum defines a joint energy over (x, c, y); prediction, intervention, and conditional interpretations are then computed as conditional probabilities by composing these energies, with gradient-based inference over relaxed concept and class variables.
Key results
- On CUB, ECBM reaches 71.3% overall concept accuracy versus 39.6% for the best baseline CEM, with class accuracy 81.2% vs 79.6% for CEM.
- On CelebA, ECBM achieves 34.3% class accuracy versus 33.0% for CEM and 47.8% overall concept accuracy versus 45.7%.
- Average L1 error between estimated and ground-truth conditional concept probabilities is 0.0033, 0.0096, and 0.0017 for the three interpretation types on CUB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace CBM/CEM with ECBM | concept prediction on CUB · CUB | Overall concept accuracy: 0.396 { $\pm$ 0.052} -> 0.713 { $\pm$ 0.009} | Helped |
| replace CEM with ECBM | class prediction · CUB | Class accuracy: 0.796 { $\pm$ 0.004} -> 0.812 { $\pm$ 0.006} | Helped |
| replace ProbCBM with ECBM | concept prediction · CelebA | Overall concept accuracy: { 0.473} { $\pm$ 0.001} -> 0.478 { $\pm$ 0.000} | Helped |
| replace CEM with ECBM | class prediction · CelebA | Class accuracy: 0.330 { $\pm$ 0.003} -> 0.343 { $\pm$ 0.000} | Helped |
| replace CBM with ECBM | concept prediction · AWA2 | Overall concept accuracy: 0.803 { $\pm$ 0.023} -> 0.854 { $\pm$ 0.000} | Helped |
| replace CEM with ECBM | class prediction · AWA2 | Class accuracy: 0.908 { $\pm$ 0.002} -> 0.912 { $\pm$ 0.000} | Helped |
| use ECBM instead of Joint (CBM) | TravelingBirds under background shift · TravelingBirds (background shift) | Class accuracy: 0.518 -> 0.584 | Helped |
| use only the x-y energy branch rather than the full ECBM (ablation) | ECBM class prediction · CUB | Class accuracy: 0.825 -> 0.812 | Helped |
| use only the x-y energy branch rather than the full ECBM (ablation) | ECBM class prediction · CelebA | Class accuracy: 0.265 -> 0.343 | Hurt |
| use only the x-y energy branch rather than the full ECBM (ablation) | ECBM class prediction · AWA2 | Class accuracy: 0.909 -> 0.912 | Hurt |
| use only the joint x-c-y energy branch rather than the full ECBM (ablation) | ECBM overall concept prediction · CUB | Overall concept accuracy: 0.680 -> 0.713 | Hurt |
| use only the joint x-c-y energy branch rather than the full ECBM (ablation) | ECBM overall concept prediction · AWA2 | Overall concept accuracy: 0.864 -> 0.854 | Helped |
Surprising findings
- ECBM underperforms vanilla CBM and CEM (RandInt) in class accuracy under concept intervention ratios, which the authors attribute to strict bottlenecks in those models.
- The single-branch ECBM (x-y only) reaches 0.909 class accuracy on AWA2 and 0.265 on CelebA, showing the global energy term drives much of the CelebA gain.
Limitations
- Global energy term requires negative sampling over concept combinations for computational efficiency, since the full space is exponential in K.
- Class accuracy under intervention lags strict-bottleneck models such as CBM and CEM with RandInt.
- Robustness claims are described as preliminary and supported by a single TravelingBirds experiment.
Experimental setup
Datasets: CUB AWA2 CelebA TravelingBirds
Benchmarks: CUB CelebA AWA2 TravelingBirds
Metrics: concept accuracy overall concept accuracy class accuracy
Baselines: CBM CEM PCBM ProbCBM Black-box
Builds on: concept bottleneck models concept embedding models probabilistic concept bottleneck models energy-based models
Closest papers by meaning
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · similarity 0.92)
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.91)
- A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · similarity 0.91)
- Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations (NeurIPS 2026 · similarity 0.91)
- VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · similarity 0.91)
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.91)
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.