Silicon Photonic Neural Networks

Abstract

With increased processing need for high bandwidth, ultra-fast, low cost, and efficient communications systems, photonics are becoming a progressively more attractive alternative to electronics. Compact size and existing infrastructure allow for easy integration of silicon photonics in VLSI systems. Due to similarities in dynamical behavior, photonics lends itself naturally to neuromorphic computing; this thesis explores using silicon photonics to create analog neural networks. We successfully demonstrate a two node recurrent photonic neural network with Hopf bifurcations induced by weight control via MRR filters. This first demonstration of such dynamics represents a giant leap towards network-based models of physical computing with integrated silicon photonics.