Optical architectures for neuron circuits

A weight-quantized learning model which is suitable for optical implementation of neural networks is proposed. This model permits reducing the required quantized levels down to two (even for the backpropagation models) by integrating the weight changes in a continuous-valued second memory. Several optical neurodevices based on AlGaAs/GaAs systems are described. They include optical neurochips and an artificial retina. Emphasis is placed on the optical learning chip with variable synaptic interconnections, which acquires the knowledge from the real world.<
>