Pré-Publication, Document De Travail Année : 2025

KL-Guided Concept-Based Learning for Explainable Classification

Résumé

Concept Bottleneck Models (CBMs) offer an interpretable framework for deep learning by introducing human-understandable concepts as intermediaries between input data and class predictions. However, the unconstrained learning of concept-to-class mappings often results in unfaithful explanations and shortcut reliance. In this work, we present KL-CBM, a novel CBM architecture that integrates a transparent, probabilistic module with a flexible dense classifier. The transparent head estimates class probabilities based on the alignment between predicted concepts and empirical class-specific profiles. Simultaneously, a dense classifier following the concept layer is softly aligned to the transparent head using Kullback-Leibler (KL) divergence. This encourages predictive consistency with interpretable modeling without sacrificing model capacity. We evaluate KL-CBM on two real-world conceptannotated datasets-AwA2 and aPY-and demonstrate that it achieves strong classification performance and competitive concept prediction accuracy. Our results highlight that KL-CBM maintains semantic alignment between concepts and classes, exhibits robustness to concept-level perturbations, and yields faithful concept-based explanations. Overall, KL-CBM advances the trade-off between interpretability and accuracy, offering a principled and flexible alternative to conventional CBMs.

Fichier principal
Vignette du fichier
KL-CBM.pdf (3.1 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05122075 , version 1 (20-06-2025)

Licence

Identifiants

  • HAL Id : hal-05122075 , version 1

Citer

Rim El Cheikh, Issam Falih, Engelbert Mephu Nguifo. KL-Guided Concept-Based Learning for Explainable Classification. 2025. ⟨hal-05122075⟩
167 Consultations
108 Téléchargements

Partager

  • More