Emergent Learning in Analog Electronic Networks

This colloquium talk by Samuel Dillavou from the University of Pennsylvania explores the groundbreaking concept of 'physical learning machines' through analog electronic networks. Discover how self-adjusting resistor networks can perform machine learning tasks like regression and classification as emergent properties of physics and local rules, rather than through conventional digital processing. Learn about the fascinating competition between physical realities (such as imperfect components) and learning processes that create unexpected dynamics. The presentation reveals how these systems exhibit double descent—a subtle transition typically only observed in pristine simulations—despite their physical constraints. Built using standard electronic components, this research establishes a foundation for studying the intersection of learning and physical processes at scale, with significant implications for developing energy-efficient AI hardware.