Modern neural networks are algorithms that acquire their functionalities by computer "learning", i.e., a computational optimization process based on gradient descent. They are algorithms arrived at by simulated self-organization: they not only designed, they are grown. The quest of my research group at Yale is to apply this same approach to physical processes. Instead of optimizing the digital weights of neural network algorithms, we optimize physical variables -- fields, currents, refractive indices, even quantum fluctuations. In this talk, I will provide an overview of a few of our ongoing efforts on this journey: building "physical" neural networks that perform analog computations at the quantum limit, training laser pulses to coax light out of molecules, and designing brain-scale analog artificial intelligence.
(This event will not have a livestream. However a recording will be made available)
