BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:João Sacramento (ETH Zürich) DTSTART:20221110T170000Z DTEND:20221110T180000Z DTSTAMP:20260423T003233Z UID:MPML/91 DESCRIPTION:Title: Th e least-control principle for learning at equilibrium\nby João Sacram ento (ETH Zürich) as part of Mathematics\, Physics and Machine Learning ( IST\, Lisbon)\n\n\nAbstract\nA large number of models of interest in both neuroscience and machine learning can be expressed as dynamical systems at equilibrium. This class of systems includes deep neural networks\, equili brium recurrent neural networks\, and meta-learning. In this talk I will p resent a new principle for learning equilibria with a temporally - and spa tially - local rule. Our principle casts learning as a least-control probl em\, where we first introduce an optimal controller to lead the system tow ards a solution state\, and then define learning as reducing the amount of control needed to reach such a state. We show that incorporating learning signals within a dynamics as an optimal control enables transmitting acti vity-dependent credit assignment information\, avoids storing intermediate states in memory\, and does not rely on infinitesimal learning signals. I n practice\, our principle leads to strong performance matching that of le ading gradient-based learning methods when applied to an array of benchmar king experiments. Our results shed light on how the brain might learn and offer new ways of approaching a broad class of machine learning problems.\ n LOCATION:https://researchseminars.org/talk/MPML/91/ END:VEVENT END:VCALENDAR