Pré-Publication, Document De Travail Année : 2024

Deep-DAP: A Self-Explainable Attribute-Based Deep Neural Network

Résumé

With the growing integration of deep learning models into everyday applications, ensuring their decisions are interpretable and trustworthy is critical. The eXplainable AI (XAI) field addresses this need by giving insights into the model behavior. Concept-based XAI methods, which link predictions to human-understandable attributes, show promise but are often post-hoc, resulting in explanations that lack robust connections to the model predictions. Attribute-based classifiers are an alternative that integrates explanatory elements into the model’s design. However, they often struggle to clearly demonstrate how attributes can be effectively leveraged to explain predictions. This paper introduces Deep-DAP, a novel architecture where the classification output is directly influenced by attribute scores, which constitute intuitive and transparent explanations. We evaluate Deep-DAP, showcasing competitive performance in both classification and attribute prediction. We also assess its explainability through sensitivity and similarity analyses, revealing Deep-DAP’s capability to provide faithful explanations alongside predictions, establishing it as a valuable tool for Explainable AI.

Fichier principal
Vignette du fichier
deepdap_HAL.pdf (2.44 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04812809 , version 1 (01-12-2024)
hal-04812809 , version 2 (17-12-2024)

Licence

Identifiants

  • HAL Id : hal-04812809 , version 2

Citer

Rim El Cheikh, Issam Falih, Engelbert Mephu Nguifo. Deep-DAP: A Self-Explainable Attribute-Based Deep Neural Network. 2024. ⟨hal-04812809v2⟩
242 Consultations
302 Téléchargements

Partager

  • More