A SMART-PIXEL OPTICAL NEURAL-NETWORK DESIGN USING CUSTOMIZED ERROR PROPAGATION

Smart-pixel technology provides an efficient method of implementing optoelectronic error-backpropagation neural networks. Quasi-linear detector/modulator circuits are required to provide the interface between optics and electronics. A design to perform the learning algorithm used by the S-SEED network is simulated using 1 mu m CMOS electronic and FET-SEED circuits. Simulations of the S-SEED neural. network demonstrate the existence of 'windows' of optimal learning behaviour with respect to the neuron contrast ratio.