PhD defense of Jean VASSOYAN

Jean VASSOYAN

Learning to teach humans and large language models

Abstract 

In classical machine learning, training data is typically assumed to stem from random sampling over which the experimenter has no control. Nevertheless, certain paradigms, such as active learning, specifically address situations where one can intervene in data selection to optimize model training. More broadly, the data selection problem can be formalized through the "teacher-student" paradigm, where the teacher is the data selection mechanism and the student is the agent to be trained. This thesis focuses on the task of the "teacher", framed as an optimization problem. We study two distinct special cases: one where the student is a human learner, and another where it is a "black-box" statistical learning model.

The first axis, centered on the human learner, falls within the field of adaptive learning, a highly active area in educational technology. We propose a new educational resource recommendation engine which, unlike classical approaches, is data-efficient, scalable, free of any a priori model of human learning, and trained end-to-end to maximize the learner's "progress". It achieves this through a keyword-based modeling of the learner's knowledge.

The second axis, focusing on the "machine" learner, relates to a family of statistical learning fields, including active, curriculum, and reinforcement learning. We introduce a unifying formal framework that allows us to bridge and compare all these domains. Finally, we study the application of this paradigm to large language models, which have become natural candidates for this type of approach. Indeed, as high-quality human textual data is nearing exhaustion, imitation learning alone is starting to show its limits. This suggests that to further improve these models, it is necessary to shift toward paradigms unconstrained by human data, based on open-ended "self-teaching" dynamics.

Key-words

Machine learning, Reinforcement learning, Adaptive learning, Teacher-student, Large language model, Graph neural networks,

PhD supervisors

• Nicolas VAYATIS, Professor, Centre Borelli, ENS Paris-Saclay (Supervisor)
• Jill-Jênn VIE

Defense committee

• Pierre-Yves OUDEYER, Directeur de recherche, Centre Inria de l'université de Bordeaux, Rapporteur
• Min CHI, Full professor, North Carolina State University, Rapporteur
• Olivier SIGAUD, Professeur des universités, Sorbonne Université, Examinateur
• Claire VERNADE, Full professor, University of Technology Nuremberg, Examinateur
• Guillaume CHARPIAT, Chargé de recherche, Centre Inria de l'Université Paris-Saclay, Examinateur

Publications