PhD defense of Antoine DE MATHELIN

Antoine De Mathelin

Towards reliable machine learning under domain shift and costly labeling, with applications to engineering design.

Abstract

In engineering design, using machine learning models to find innovative products poses major challenges. The effectiveness of machine learning models has been demonstrated when trained and used on large datasets of independently identically distributed observations. However, in the engineering design context, models are often deployed on shifted distributions, with few labeled data available. Moreover, the reliability of the model is strongly required as trusting wrong predictions can lead to dramatic consequences. This thesis tackles the challenge of providing a reliable machine learning model under the main engineering design constraints: domain shift and costly labeling. By leveraging novel contributions from domain adaptation, active learning and uncertainty quantification techniques, we propose a generic approach towards this goal. Moreover, the contributions of this thesis to the three aforementioned thematics are impactful beyond the context of engineering design. They allow for achieving similar or better performances with less data and reduced computation time compared to standard approaches. Additionally, the thesis delivers accessible and user-friendly tools through a domain adaptation and transfer learning library called Adapt.

Supervision

• Mathilde MOUGEOT
• Nicolas VAYATIS
• François DEHEEGER, Senior Fellow AI & data science, Michelin R&D.

Jury

• Amaury HABRARD, Professeur, Université Jean Monnet de Saint-Etienne, Rapporteur & Examinateur
• Jean-Michel LOUBES, Professeur, Université de Toulouse, Rapporteur & Examinateur
• Josselin GARNIER, Professeur, Ecole Polytechnique, Examinateur
• Céline HUDELOT, Professeure, Ecole CentraleSupelec, Examinatrice
• Alexandre GRAMFORT, Senior Research Scientist, Meta, Examinateur
• Benjamin GUEDJ, Associate Professor, University College London, Examinateur