ALP Seminar: Circuits that learn on their own

Artificial neural networks are digital and learn by global optimization, using vast amounts of memory, processor time, and energy. By contrast, neural networks in the brain are analogue and learn by local rules. I will discuss our laboratory circuits of self-adjusting nonlinear resistors that update their resistances by local rules, enabling the networks to perform tasks such as classification and regression. Learning emerges as a collective phenomenon from identical electrical components updating via identical local rules, without a processor or external memory, and using very little power.