R&D Unplugged #19 - Learning to teach humans and neural networks, with Jean Vassoyan
In this talk, Jean Vassoyan presents his work on a relatively broad machine learning problem: the design of “teacher” algorithms built to optimise the progress of “student” agents.
While purposely abstract, this framework unifies a wide range of research fields-including Curriculum Learning, Active Learning, and Reinforcement Learning. When applied to humans, this concept translates to Adaptive Learning: the creation of intelligent tutors capable of personalizing education for each individual.
Although these domains may seem distinct, their objectives can be formalised using a common mathematical language. This unified view allows us to step back, compare these methods effectively, and try to bridge the gap between machine and human instruction.
Ultimately, this presentation addresses a core question: How do we define the role of a teacher and formalize it from a quantitative perspective?
Event open to all
→ register here
More about the speaker
Jean Vassoyan is a graduate of Télécom Paris and the MVA Master's program. He is currently pursuing a CIFRE PhD at Centre Borelli in partnership with onepoint. His early research focused on Reinforcement Learning applied to Intelligent Tutoring Systems. More recently, he has shifted his attention to Large Language Model (LLM) post-training, specifically exploring Reinforcement Learning with Verifiable Rewards and methods to optimize training tasks.
Organised by the Learning Transitions Research Unit, R&D Unplugged series - the Talk and the Podcast - focuses on the purpose of Research, its key data, and its tangible impact on the real-world. Sign up to receive news (in English) from our R&D Unplugged series: more info here.
