🤖 AI Summary
This work addresses the problem that Concept Activation Vectors (CAVs) are contaminated by irrelevant interference signals due to their reliance on linear classifier optimization. To mitigate this, we propose a pattern-driven, unsupervised CAV construction method that abandons pseudo-labeling and classification objectives. Instead, it directly extracts pure concept directions from latent representations via feature pattern decomposition and concept-signal focusing—applicable to mainstream vision models including VGG, ResNet, and ViT. Our key contribution is the first identification and explicit mitigation of classification-induced bias, establishing the first supervision-free CAV paradigm. Experiments on Pediatric Bone Age, ISIC2019, and FunnyBirds demonstrate that our CAVs better align with the true geometric structure of human-interpretable concepts, significantly improving robustness in concept-sensitivity detection and enhancing correction of model shortcut behaviors.
📝 Abstract
With a growing interest in understanding neural network prediction strategies, Concept Activation Vectors (CAVs) have emerged as a popular tool for modeling human-understandable concepts in the latent space. Commonly, CAVs are computed by leveraging linear classifiers optimizing the separability of latent representations of samples with and without a given concept. However, in this paper we show that such a separability-oriented computation leads to solutions, which may diverge from the actual goal of precisely modeling the concept direction. This discrepancy can be attributed to the significant influence of distractor directions, i.e., signals unrelated to the concept, which are picked up by filters (i.e., weights) of linear models to optimize class-separability. To address this, we introduce pattern-based CAVs, solely focussing on concept signals, thereby providing more accurate concept directions. We evaluate various CAV methods in terms of their alignment with the true concept direction and their impact on CAV applications, including concept sensitivity testing and model correction for shortcut behavior caused by data artifacts. We demonstrate the benefits of pattern-based CAVs using the Pediatric Bone Age, ISIC2019, and FunnyBirds datasets with VGG, ResNet, ReXNet, EfficientNet, and Vision Transformer as model architectures.
