Abstract

The complexities of implementing neural network systems stem from the requirement that each neuron can receive excitation from many inputs (1-1000, or more) and each input must be multiplied by a weight Conventional analog and digital electronic hardware implementations of neural architectures often use much of the available hardware to implement the calculation of the product of the weights and inputs, and have to resort to a time-multiplexing scheme (which allows sharing of the multiplier hardware) to implement networks with more than a few thousand neurons in the system. This problem can be overcome by using stochastic computing techniques. Therefore, this paper details the results of an investigation of the implementation of the functional components of a stochastic bit stream neuron in optic/optoelectronic hardware. This approach offers several advantages.

© 1995 Optical Society of America