A new optoelectronic technology for the implementation of neural network architectures is being developed at Lincoln Laboratory. The new technology is based on a multiple-quantum- well (MQW) device called the monolithic optoelectronic transistor (MOET). MOET is a true optical transistor; it enables the switching of one optical signal with a much weaker one. The terminal characteristics are ideal for implementing a neuron: abrupt or sigmoidal thresholds, saturated ON and OFF states, and high fan-out. The device will be initially demonstrated for implementation of early visual processing networks. The baseline network is the CORT-X model, a multiple spatial-scale, feedforward network for boundary segmentation of noisy binary images. MOET implementation is possible with slight modifications to the CORT-X architecture. Simulations of the hardware implemented network are carried out and compared with the performance of the original model. As a function of input image carrier-to-noise ratio (CNR), performance is evaluated with respect to deviations from ideal response along two dimensions: (1) contrast-ratio and (2) nonuniformity. Assuming ideal hardware response, the modified CORT-X architecture performs better than the original model. Moderate contrast does not significantly degrade network performance, while nonuniformities as small as 10% degrade performance even for high CNR.
展开▼
