Year: 2025
Type: preprint
Abstract: Demand for interpretable, controllable models is growing as machine learning enters high-stakes, human-centered applications. This survey fills a missing intersection in the literature by systematically reviewing ad-hoc, concept-based methods—approaches that embed human-interpretable concepts directly into the model architecture or learning process so concepts functionally mediate prediction. We synthesize the field into six families: explicit concept layers, latent concept embeddings, probabilistic representations, attention-based mechanisms, prototype-based evidence, and symbolic/logic reasoning. We provide comparison tables and a practitioner-oriented analysis along axes such as concept learning strategies, degree of intervenability, and explanation formats, offering guidance on method selection under different constraints. We also examine broader issues in concept set design, the role of language/vision-language models, cross-modal applicability, and evaluation protocols for models and explanations. We conclude with open challenges and research directions toward transparent, editable AI systems aligned with human goals.
Source: HAL
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Agence Nationale de la Recherche
OpenAlex ID: W7140023801
