An Optical Neural Network Architecture based on Highly Parallelized WDM-Multiplier-Accumulator

Abstract

Future applications such as anomaly detection in a network and autonomous driving require extremely low, submicrosecond latency processing in pattern classification. Towards the realization of such an ultra-fast inference processing, this paper proposes an optical neural network architecture which can classify anomaly patterns at sub-nanosecond latency. The architecture fully exploits optical parallelism of lights using wavelength division multiplexing (WDM) in vector-matrix multiplication. It also exploits a linear optics with passive nanophotonic devices such as microring resonators, optical combiners, and passive couplers, which make it possible to construct low power and ultra-low latency optical neural networks. Optoelectronic circuit simulation using optical circuit implementation of multi-layer perceptron (MLP) demonstrates sub-nanosecond processing of optical neural network.