A multi-layer-perceptron neural network hardware based on 3D massively parallel optoelectronic circuits

Abstract

A digital neural network architecture is presented which is based on three-dimensional massively parallel optoelectronic circuits. A suitable optical interconnect system and the structure of the required electronic circuits is specified. For this system general formulas for the performance of such a neural network architecture are determined. A parameter study using current technological limitations and timing values from electronic implementation is carried out. Based on this analysis if is shown that this novel type of neuroarchitecture that is using 3D massively parallel optoelectronic circuits shows performance rates of up to one magnitude higher than systems using digital neurochips based on fully electronic implementation.