Bonjour à tous,

Vous êtes cordialement invité.e.s à la soutenance de thèse de Hajer Mejri le jeudi 15 septembre à 14h00 EST.

Title : Toward Reasoned Deep Learning: Leveraging Causality for Interpretability and Generalization

Date: Jeudi 15 octobre à 14h00 

Salle:  Pavillon André-Aisenstadt, 2920 Ch. de la Tour, au AA-3195

Jury

Résumé:

While recent advancements in AI are reshaping daily life, work, and the global economy, there are still persistent limitations, such as explainability and generalization. In this dissertation, we explore how leveraging causality can address these challenges. For interpretability, we demonstrate that with the integration of causal models and counterfactual reasoning, one may explain the decision-making process of image based deep learning classifiers.  In the context of plant disease classification, we present “Beyond Accuracy", a causal inference evaluation framework to measure the Causal Effect (CE) of Northern Leaf Blight lesions in image classifier predictions which leverage counterfactual interventions.  For generalization, we consider the role of causal inference in removing spurious correlations while preserving the core semantic features that remain stable across varying data distributions and demonstrate with proper causality design, we achieve better performance in generalization.