BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:Optimal Learning Protocols Via Statistical Physics and Control The ory - Francesca Mignacco (Princeton University) DTSTART:20250910T143000Z DTEND:20250910T151000Z UID:TALK233296@talks.cam.ac.uk DESCRIPTION:Learning is a complex dynamical process shaped by many interco nnected decisions. Protocols that govern how to tune hyperparameters in ar tificial networks\, or how to allocate cognitive effort in biological lear ners\, can have dramatic effects on performance. Yet our theoretical under standing of optimal learning strategies remains limited\, due to the nonli near nature of learning dynamics and the high dimensionality of the learni ng space. \;\nIn this talk\, I will present a framework that combines statistical physics and control theory to identify optimal learning protoc ols in prototypical neural network models (see Refs. [1\,2]). In the high- dimensional limit\, we derive closed-form equations for a small set of ord er parameters that track stochastic gradient descent. This reduction allow s to formulate the design of learning protocols&mdash\;such as curricula\, dropout schedules\, or noise levels&mdash\;as an optimal control problem on the dynamics of the order parameters\, with the objective of minimizing the final generalization error. \;\nI will discuss applications to bo th toy models and real datasets\, showing how the resulting strategies unv eil key learning tradeoffs\, for example\, between exploiting informative directions in the data and limiting noise sensitivity\, and how these insi ghts may contribute to a principled theory of meta-learning. \;\n[1] M ignacco\, F. and Mori\, F.\, 2025. A statistical physics framework for opt imal learning. arXiv preprint arXiv:2507.07907. \;\n[2] Mori\, F.\, Ma nnelli\, S.S. and Mignacco\, F.\, Optimal Protocols for Continual Learning via Statistical Physics and Control Theory. In The Thirteenth Internation al Conference on Learning Representations. \; LOCATION:External END:VEVENT END:VCALENDAR