α-TCAV: A Unified Framework for Testing with Concept Activation Vectors
Abstract
Concept Activation Vectors (CAV) and the Testing with CAVs (TCAV) method are fundamental tools for concept-based explainability in deep learning, yet their practical utility is limited by the variability across different samples of concept and non-concept examples. We analyze this stochastic nature of CAVs, and identify a fundamental flaw in the standard TCAV score, its instability by using a discontinuous indicator function. To address this, we introduce α-TCAV, a parameterized unified framework that replaces the discontinuous indicator with a smooth function, which recovers TCAV as a limiting case. Under realistic assumptions, we characterize the score distributions and derive principled parameter choices: one matches the mean of its averaged counterpart, Multi-TCAV, with no larger variance using a single CAV; the other provides a calibrated probability that the concept positively influences the model output. Our analysis and experiments provide practical guidance and challenge the established practice of splitting the sampling budget across multiple CAVs.
Related
- Concept Modulation Models: A Unified Framework for Identifiability and Extrapolation
- A unified framework for spatially resolved cortical activation analysis
- A Unified Framework for Scenario Coverage Estimation and Representative Selection in Simulation-Based Testing of Automated Vehicles
- A Unified Complexity Framework for Quantum Property Testing
- Beyond Marginal Monitoring: Distributed Joint-Distribution Testing for Data Concept Drift in Large Scale E-Commerce Operations
BibTeX
@misc{schnoor2026αtcav,
title = {{$α$-TCAV: A Unified Framework for Testing with Concept Activation Vectors}},
author = {Ekkehard Schnoor and Jawher Said and Malik Tiomoko and Wojciech Samek and Alexander Jung},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2605.15688},
eprint = {2605.15688},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.15688},
}
RIS
TY - UNPB
TI - α-TCAV: A Unified Framework for Testing with Concept Activation Vectors
AU - Ekkehard Schnoor
AU - Jawher Said
AU - Malik Tiomoko
AU - Wojciech Samek
AU - Alexander Jung
PY - 2026
DA - 2026/10/12
PB - arXiv
DO - 10.48550/arxiv.2605.15688
UR - https://arxiv.org/abs/2605.15688
AB - Concept Activation Vectors (CAV) and the Testing with CAVs (TCAV) method are fundamental tools for concept-based explainability in deep learning, yet their practical utility is limited by the variability across different samples of concept and non-concept examples. We analyze this stochastic nature of CAVs, and identify a fundamental flaw in the standard TCAV score, its instability by using a discontinuous indicator function. To address this, we introduce α-TCAV, a parameterized unified framework that replaces the discontinuous indicator with a smooth function, which recovers TCAV as a limiting case. Under realistic assumptions, we characterize the score distributions and derive principled parameter choices: one matches the mean of its averaged counterpart, Multi-TCAV, with no larger variance using a single CAV; the other provides a calibrated probability that the concept positively influences the model output. Our analysis and experiments provide practical guidance and challenge the established practice of splitting the sampling budget across multiple CAVs.
ER -