LFI / LIP6 "Balancing fidelity and interpretability in XAI for global model understanding"

Abstract : As neural networks grow larger and more complex, they achieve strong performance but also introduce important risks. For example, a facial attribute recognition model may rely on biased correlations related to gender or race rather than on meaningful visual features, leading to unfair or misleading decisions despite high accuracy. Explainable Artificial Intelligence (XAI) aims to make these black-box models more understandable. In particular, post-hoc explanation methods identify which features influence a model’s predictions without changing the model itself. However, highly faithful explanations are often difficult for humans to interpret, while simpler explanations may not accurately reflect the model’s reasoning. In this work, we study how to balance fidelity and interpretability to improve explanations for model auditing and debugging.

Bio : Caroline Mazini Rodrigues is a postdoctoral researcher at the Institut de Recherche en Informatique et Systèmes Aléatoires (IRISA), where she develops low-complexity algorithms for video compression networks, aiming for simplicity, interpretability, and frugality in AI models. She earned her PhD from Université Gustave Eiffel, conducting research at the Laboratoire d’Informatique Gaspard-Monge (LIGM) and the Laboratoire de Recherche de l’EPITA (LRE) on explainable Artificial Intelligence (XAI), with a particular focus on understanding the reasoning processes of deep neural networks. Her research interests include making AI more transparent, frugal, and human-interpretable, while exploring the cognitive aspects of machine learning and its connections to human reasoning and learning.
Website: https://carolmazini.github.io/