Effects of Device Characteristics on the Performance of Optical Neural Networks

Abstract

An optical neuron architecture model incorporating optical devices is given. From the fabrication point of view the permissible density of neurons in the network with light Emitting Diode (LED) and Laser as transmitter are evaluated. For this model an optical feed-forward neural network is simulated and trained using the back propagation algorithm. Further, it has also been demonstrated that optoelectronic neural networks using either LED or laser, can learn and function satisfactorily even in the presence of non-ideal network characteristics such as optical cross-talk, optoelectronic device performance variations and nonlinear response of optoelectronic output devices. While cross talk up to about 60% is acceptable both during learning and recalling process, the performance of the model is practically independent of device behaviour.