A Survey Of Ad-Hoc Concept-Based Explainability
Résumé
As machine learning systems are increasingly deployed in high-stakes and human-centered domains, the need for models that are interpretable, controllable, and aligned with human reasoning has grown significantly. Concept-based explainability methods aim to address this need by grounding the behavior of deep learning models in semantically meaningful units known as concepts. In this survey, we focus on ad-hoc concept-based explainability methods, which are approaches that embed concepts directly into the architecture or learning process of a predictive model, rather than approximating them after the training. These methods support structural interpretability, enable faithful explanatory insights, and allow users to interact directly with the prediction process. We present a comprehensive taxonomy of approaches designed for ad-hoc concept-based explainability, organized by the form of concept representation and integration strategy. We provide a detailed comparative analysis across different dimensions, including concept learning strategies, degrees of intervenability, and explanation formats. In doing so, we highlight the trade-offs and design choices that shape current approaches in this field. Finally, we cover methodological considerations and challenges, and explore emerging trends-concept discovery, leveraging large language models for concept supervision, and the current framework of evaluation metrics. This survey aims to map the current landscape of ad-hoc concept-based explainability methods, offering a descriptive overview of existing approaches as a foundation for future work in the development of AI systems that are not only accurate, but also transparent, editable, and aligned with human understanding.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |