Research map / Label-efficient and robust learning / Recognition and long-tail
Hierarchical Concept-based Interpretable Models
Oscar Hill, Mateo Espinosa Zarlenga, Mateja Jamnik
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 |
|---|---|---|---|
| add Concept Splitting sub-concept labels to HiCEM training (ablation) | HiCEM · PseudoKitchens | discovered sub-concept ROC-AUC: 0.75 -> 0.88 | Improved |
| add Concept Splitting sub-concept labels to HiCEM training (ablation) | HiCEM · AwA2 | discovered sub-concept ROC-AUC: 0.73 -> 0.88 | Improved |
| add Concept Splitting sub-concept labels to HiCEM training (ablation) | HiCEM · CUB | discovered sub-concept ROC-AUC: 0.81 -> 0.85 | Improved |
| replace CEM with HiCEM after Concept Splitting | CEM · CUB | task accuracy: 0.79 -> 0.74 | Hurt |
| replace CEM with HiCEM after Concept Splitting | CEM · MNIST-ADD, AwA2, PseudoKitchens | task accuracy: 0.93 -> 0.92 (MNIST-ADD); 0.98 -> 0.98 (AwA2); 0.66 -> 0.65 (PseudoKitchens) | No clear effect |
| use Concept Splitting sub-concepts as supervision | HiCEM · ImageNet user study | % yes sub-concept related to parent: 4.0% -> 67.9% | Improved |
| use Concept Splitting image labels | HiCEM · ImageNet user study | % yes image consistent with name: 0.9% -> 54.8% | Improved |
| intervene on discovered sub-concepts in HiCEM instead of CEM with discovered concepts | HiCEM · CUB, PseudoKitchens | task accuracy after interventions | Mixed |
| add concept splitting supervision to HiCEM before interventions (ablation) | HiCEM · interventions | task accuracy | Improved |
| SAE vs clustering for Concept Splitting (ablation) | HiCEM · discovered sub-concept ROC-AUC | ROC-AUC: CUB 0.90 (clustering) vs 0.85 (SAE); AwA2 0.81 vs 0.88 | Mixed |
Closest papers by meaning
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.88)
- Hyperbolic Concept Bottleneck Models (NeurIPS 2026 · similarity 0.87)
- Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models (CVPR 2024 · similarity 0.87)
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.86)
- Understanding Inter-Concept Relationships in Concept-Based Models (ICML 2024 · similarity 0.86)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.86)
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.
