Explaining Explainability: Recommendations for Effective Use of Concept Activation Vectors

TL;DR

This paper investigates properties of Concept Activation Vectors (CAVs) used in interpretability, identifies challenges like entanglement and spatial dependency, and offers tools and recommendations to improve concept-based explanations in deep learning models.

Contribution

It introduces tools to detect properties affecting CAV interpretability, demonstrates their impact on real tasks, and proposes methods to mitigate misleading explanations.

Findings

Entanglement affects interpretability of CAVs.

Negative probe set choice influences CAV meaning.

Spatially dependent CAVs can test translation invariance.

Abstract

Concept-based explanations translate the internal representations of deep learning models into a language that humans are familiar with: concepts. One popular method for finding concepts is Concept Activation Vectors (CAVs), which are learnt using a probe dataset of concept exemplars. In this work, we investigate three properties of CAVs: (1) inconsistency across layers, (2) entanglement with other concepts, and (3) spatial dependency. Each property provides both challenges and opportunities in interpreting models. We introduce tools designed to detect the presence of these properties, provide insight into how each property can lead to misleading explanations, and provide recommendations to mitigate their impact. To demonstrate practical applications, we apply our recommendations to a melanoma classification task, showing how entanglement can lead to uninterpretable results and that the choice of negative probe set can have a substantial impact on the meaning of a CAV. Further, we show that understanding these properties can be used to our advantage. For example, we introduce spatially dependent CAVs to test if a model is translation invariant with respect to a specific concept and class. Our experiments are performed on natural images (ImageNet), skin lesions (ISIC 2019), and a new synthetic dataset, Elements. Elements is designed to capture a known ground truth relationship between concepts and classes. We release this dataset to facilitate further research in understanding and evaluating interpretability methods.