$\alpha$-TCAV: A Unified Framework for Testing with Concept Activation Vectors
The article presents a new framework called α-TCAV for testing with Concept Activation Vectors in deep learning. It addresses the limitations of the existing TCAV method by introducing a smooth function to improve statistical stability. The authors provide guidance on optimizing the framework for better computational efficiency and accuracy in measuring concept influence.
- ▪Concept Activation Vectors (CAVs) are essential for explainability in deep learning but suffer from statistical instability.
- ▪The α-TCAV framework replaces the discontinuous indicator function in TCAV with a parameterized smooth function.
- ▪The authors recommend allocating the full sampling budget to a single CAV instead of distributing it across multiple CAVs.
Opening excerpt (first ~120 words) tap to expand
Statistics > Machine Learning arXiv:2605.15688 (stat) [Submitted on 15 May 2026] Title:$α$-TCAV: A Unified Framework for Testing with Concept Activation Vectors Authors:Ekkehard Schnoor, Jawher Said, Malik Tiomoko, Wojciech Samek, Alexander Jung View a PDF of the paper titled $\alpha$-TCAV: A Unified Framework for Testing with Concept Activation Vectors, by Ekkehard Schnoor and 4 other authors View PDF Abstract:Concept Activation Vectors (CAVs) are a fundamental tool for concept-based explainability in deep learning, yet their practical utility is limited by statistical instability. We analyze the stochastic nature of CAVs and the Testing with CAVs (TCAV) method, deriving the distributions of major CAV classes including PatternCAV, FastCAV, and ridge regression-based CAVs.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
