Research map / Label-efficient and robust learning / Recognition and long-tail
Concept Bottleneck Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use a concept bottleneck rather than a standard end-to-end model | OAI KLG prediction · OAI test | KLG RMSE: 0.441$\pm$ 0.006 -> 0.418$\pm$ 0.004 | Helped |
| use a concept bottleneck rather than a standard end-to-end model | CUB species classification · CUB test | 0-1 error: Standard: 0.175$\pm$0.008; Joint: 0.199$\pm$0.006; Independent: 0.240$\pm$0.012 | Mixed |
| predict concepts with a trained bottleneck instead of probing standard-model activations | OAI concept prediction · OAI test | Concept RMSE: 0.680$\pm$0.038 -> 0.529$\pm$0.004 | Helped |
| predict concepts with a trained bottleneck instead of probing standard-model activations | CUB concept prediction · CUB test | Concept error: 0.093$\pm$0.004 -> 0.031$\pm$0.000 | Helped |
| use concept bottlenecks rather than a standard end-to-end model | Bird species classifier · TravelingBirds background shift | Task error: 0.627$\pm$0.013 -> 0.482$\pm$0.008 | Helped |
| add test-time corrections to predicted concepts | Concept bottleneck models · OAI test | KLG RMSE: from ${>}0.4 to ${\approx}0.3 | Helped |
| add test-time corrections to predicted concepts | Concept bottleneck models · CUB test | Classification accuracy | Helped |
| add a bottleneck layer to the standard model (ablation) | Standard end-to-end model · OAI and CUB test | Task error: OAI: 0.441$\pm$ 0.006 -> 0.443 $\pm$ 0.008; CUB: 0.175$\pm$0.008 -> 0.173$\pm$0.003 | No clear effect |
| add an auxiliary concept-prediction loss (ablation) | Standard end-to-end model · OAI and CUB test | Task error: OAI: 0.441$\pm$ 0.006 -> 0.425$\pm$ 0.010; CUB: 0.175$\pm$0.008 -> 0.162$\pm$0.002 | Helped |
| set joint-training lambda to 0.01, prioritizing task fit over concept fit (ablation) | Joint bottleneck · OAI test-time intervention | Test error | Hurt |
| replace the 3-layer MLP with a single linear layer for concept-to-target prediction (ablation) | OAI concept bottleneck · OAI test-time intervention | Task accuracy | Hurt |
| connect the final predictor to concept probabilities instead of logits (ablation) | Joint bottleneck · CUB test | 0-1 error: 0.199 -> 0.224 | Hurt |
Closest papers by meaning
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.84)
- Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable? (NeurIPS 2024 · similarity 0.84)
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · similarity 0.83)
- Bayesian Concept Bottleneck Models with LLM Priors (NeurIPS 2025 · similarity 0.82)
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.82)
- A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis (NeurIPS 2024 · 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.