α-TCAV: A Unified Framework for Testing with Concept Activation Vectors
A recent publication on arXiv (2605.15688) presents α-TCAV, a comprehensive framework aimed at enhancing concept-based explainability within deep learning. The researchers investigate the stochastic characteristics of Concept Activation Vectors (CAVs) and the Testing with CAVs (TCAV) approach, deriving distributions for key CAV categories such as PatternCAV, FastCAV, and ridge regression-based CAVs. They uncover a significant issue with the traditional TCAV score, noting that its dependence on a discontinuous indicator function leads to persistent variance in crucial areas. By substituting the indicator with a smooth, parameterized function, α-TCAV offers a cohesive probabilistic model that encompasses both TCAV and Multi-TCAV. The study also delineates the distributions of sensitivity scores and various TCAV types, revealing that existing state-of-the-art selections lack theoretical support.
Key facts
- arXiv paper 2605.15688 introduces α-TCAV
- α-TCAV is a generalized framework for concept-based explainability
- Analyzes stochastic nature of CAVs and TCAV method
- Derives distributions for PatternCAV, FastCAV, and ridge regression-based CAVs
- Identifies flaw in standard TCAV score: discontinuous indicator function causes non-decaying variance
- α-TCAV replaces indicator with parameterized smooth function
- Unified probabilistic formulation subsumes TCAV and Multi-TCAV
- Shows state-of-the-art choices lack theoretical justification
Entities
Institutions
- arXiv