CREDENCE Framework Decomposes Concept Uncertainty in AI Models
A novel framework named CREDENCE (Credal Ensemble Concept Estimation) has been developed by researchers for Concept Bottleneck Models (CBMs). This innovative approach distinctly separates concept uncertainty into two types: epistemic and aleatoric. In contrast to traditional CBMs that provide point probabilities merging reducible model underspecification with irreducible input ambiguity, CREDENCE characterizes each concept through a credal prediction, represented as a probability interval. Epistemic uncertainty arises from varying opinions among different concept heads, while aleatoric uncertainty is calculated using a specific ambiguity output that aligns with annotator disagreement when applicable. This separation allows for more informed decision-making, including automating low-uncertainty instances and directing high-aleatoric cases for human evaluation. The findings are presented in arXiv preprint 2604.24170.
Key facts
- CREDENCE stands for Credal Ensemble Concept Estimation.
- It is a framework for Concept Bottleneck Models (CBMs).
- It decomposes concept uncertainty into epistemic and aleatoric components.
- Epistemic uncertainty is derived from disagreement across diverse concept heads.
- Aleatoric uncertainty is estimated via a dedicated ambiguity output.
- The framework supports prescriptive decisions like automation and data collection.
- The preprint is available on arXiv with ID 2604.24170.
- The work addresses conflation of uncertainty types in standard CBMs.
Entities
Institutions
- arXiv