Research map / Label-efficient and robust learning / Recognition and long-tail
Leakage and Interpretability in Concept-Based Models
Enrico Parisini, Tapabrata Chakraborti, Chris Harbron, Ben D. MacArthur, Christopher R.S. Banerji
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 |
|---|---|---|---|
| use CTL instead of OIS as leakage indicator (ablation) | leakage scores · TabularToy(0.25) | Pearson r with S_int: 0.75 vs 0.49 | Improved |
| use CTL instead of OIS as leakage indicator (ablation) | leakage scores · HAM10K | Pearson r with S_int: 0.49 vs -0.06 | Improved |
| use CTL instead of OIS as leakage indicator (ablation) | leakage scores · dSprites(4) | Pearson r with S_int: 0.67 vs 0.55 | Improved |
| use ICL instead of OIS (ablation) | leakage scores · 3dshapes(0) | Pearson r with S_int: 0.86 vs 0.82 | Improved |
| increase concept supervision lambda in soft/logit CBMs (ablation) | soft and logit CBMs · TabularToy, dSprites, 3dshapes, CUB, HAM10K | CTL and ICL leakage: leakage decreases as concept supervision increases | Improved |
| increase lambda beyond about 10 (ablation) | soft and logit CBMs · TabularToy(0.25) | task accuracy: task accuracy above 90% in only 3 out of 5 folds for lambda = 10, 20, 50 | Hurt |
| use logit instead of soft probability concept encoding for binary concepts (ablation) | CBMs · several datasets | minimal attainable leakage: minimal leakage higher for logit CBMs than soft CBMs | Hurt |
| complete the annotated concept set (vs incomplete) (ablation) | soft CBMs · 15 model classes | CTL/ICL leakage: more leakage in 14/15 model classes with incomplete concepts | Improved |
| use a flexible final head instead of a misspecified linear head (ablation) | soft CBMs · TabularToy, dSprites, 3dshapes | CTL/ICL leakage: misspecified head causes higher leakage in 7/9 model classes | Improved |
| incomplete concept set lowers reference task accuracy (ablation) | final head · 3dshapes(0) | y_acc: 1.000 -> 0.655 | Hurt |
| non-linear task with linear head (ablation) | soft CBM linear head · TabularToy(0.25) | y_acc: 1.000 -> 0.687 | Hurt |
| increase lambda and p_int in CEMs (ablation) | CEMs · TabularToy, CUB and others | interconcept leakage (vectors): interconcept information in vectors grows with lambda and p_int | Hurt |
Closest papers by meaning
- Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts (ICML 2025 · similarity 0.88)
- A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · similarity 0.86)
- An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations (NeurIPS 2025 · similarity 0.86)
- Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.86)
- Understanding Inter-Concept Relationships in Concept-Based Models (ICML 2024 · similarity 0.85)
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.84)
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.
