Research map / Label-efficient and robust learning / Recognition and long-tail
Understanding Inter-Concept Relationships in Concept-Based Models
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.
Analyzes whether concept-based models (TCAV, CEM, Concept2Vec) learn concept representations that capture inter-concept relationships, introduces stability/robustness/responsiveness/faithfulness metrics, and proposes basis-aided concept intervention that imputes unintervened concepts from similar intervened ones.
Problem
Concept-based models typically predict concepts independently, and it is unclear whether their learned concept representations capture known inter-concept relationships in a stable, robust, and faithful way.
Approach
Define metrics (stability, robustness, responsiveness, faithfulness) over concept bases compared via overlap of nearest-neighbour concepts, and use the learned concept similarities to impute unintervened concepts during test-time concept intervention, with theoretical bounds linking co-occurrence estimation to intervention accuracy.
Key results
- On CUB, Label bases achieve faithfulness 0.73 and stability 1.00 while CEM reaches 0.06 faithfulness and 0.03 stability and TCAV 0.08 faithfulness and 0.01 stability.
- On coloured MNIST, CEM bases recover digit-colour concept agreement only 87% ± 19%, versus 100% for TCAV, Concept2Vec and Label bases.
- Basis-aided intervention with Label bases improves CUB intervention accuracy for 20%-80% of concept groups intervened, while poorly constructed TCAV/CEM bases hurt accuracy by up to 10% on CUB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use label-basis representations for concept intervention | CEM · CUB | concept intervention accuracy | Helped |
| Use TCAV or CEM concept bases instead of label bases | CEM concept intervention · CUB | concept intervention accuracy: up to 10% | Hurt |
| Use label-basis representations for concept intervention | CEM · MNIST | concept intervention accuracy | Helped |
| Use label-basis representations for concept intervention when more than 20% of concepts are known | CEM · dSprites | concept intervention accuracy | Helped |
| Use concept bases for intervention | CEM · CheXpert | concept intervention accuracy | No clear effect |
| Use label bases with fewer training epochs instead of training longer without them | CEM concept intervention · CUB | concept intervention accuracy: 50 epochs with label bases ... 100 epochs without label bases | Helped |
| Increase correlation between number and colour concepts | CEM · Coloured MNIST | concept, task, and intervention accuracy: 0% to 100% | Helped |
| Use CEM concept representations rather than label co-occurrence vectors | concept agreement · Coloured MNIST | concept agreement: 100% ± 0% -> 87% ± 19% | Hurt |
| Increase concept-flip rate to 10% (ablation) | Concept2Vec basis · CUB | robustness: decreases by 0.2 | Hurt |
| Vary t from 1 to 3 (ablation) | Label basis faithfulness metric · CUB | faithfulness | Helped |
| Use Manhattan distance instead of cosine or Euclidean distance (ablation) | concept basis construction · Multiple datasets | basis similarity | Hurt |
| Increase r after q is at least 10 (ablation) | basis-aided concept intervention · Multiple datasets | intervention performance | No clear effect |
Surprising findings
- CEM and TCAV concept bases fail to recover known digit-colour relationships on coloured MNIST and show very low stability on CUB.
- Training with fewer epochs plus label bases outperforms training longer without them, suggesting gains from concept bases cannot be replicated by extra compute.
Limitations
- Evaluation limited to image datasets; text and other modalities not studied.
- Only CEM and TCAV (and Concept2Vec) analysed; newer concept-based models not evaluated.
- CheXpert interventions showed minimal impact, so the method was not validated there.
Experimental setup
Datasets: Coloured MNIST dSprites CUB CheXpert
Benchmarks: Coloured MNIST dSprites CUB CheXpert
Metrics: stability robustness responsiveness faithfulness concept agreement concept intervention accuracy Oracle Impurity Score
Baselines: TCAV CEM Concept2Vec Label basis
Builds on: concept bottleneck models concept embedding models TCAV word2vec
Closest papers by meaning
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.87)
- Walking the Web of Concept-Class Relationships in Incrementally Trained Interpretable Models (AAAI 2025 · similarity 0.86)
- Concept frustration: Aligning human concepts and machine representations (NeurIPS 2026 · similarity 0.85)
- Interpretable Concept-Based Memory Reasoning (NeurIPS 2024 · similarity 0.85)
- Causally Reliable Concept Bottleneck Models (NeurIPS 2025 · similarity 0.85)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · similarity 0.85)
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.