Séminaire Image: Balancing fidelity and interpretability in XAI for global model understanding and frugal AI design par Caroline Mazini Rodrigues

Nous aurons le plaisir d’écouter Caroline Mazini Rodrigues, Postdoctorante à l’IRISA (Rennes).
Elle donnera un séminaire IMAGE le jeudi 18 juin 2026 à 14h en salle de séminaire F-200

Titre: Balancing fidelity and interpretability in XAI for global model understanding and frugal AI design

Résumé:

As neural networks grow in complexity, understanding their decision-making processes has become important for ensuring transparency, fairness, and trustworthiness. Explainable Artificial Intelligence (XAI) addresses this challenge by identifying the features that influence model predictions. However, highly faithful explanations are often difficult for humans to interpret, while simpler explanations may fail to accurately reflect the model’s reasoning. This talk presents work on balancing the faithfulness-interpretability trade-off in post-hoc explanation methods, with applications to bias understanding in vision models. We then explore another research direction at the intersection of XAI and Frugal AI, examining how explanation-based insights can guide the design of more efficient and resource-aware AI models. This line of work positions interpretability not only as a diagnostic tool, but as a mechanism for improving computational efficiency and responsible AI.