A large hybrid electro-optic computer for neural network applications

We report on the design, construction and operation of a hybrid electro-optic computer intended for neural network applications. We have recently configured this system to implement what we believe is the largest fully interconnected neural network built to date. Our approach uses photoconduction in amorphous silicon films to achieve a fully programmable synapse. The 14,400 synapses in this network possess full analog depth and can be dynamically programmed over three decades of connection strength. The conductivity at each synapse is independently determined by the light flux it receives. We describe the use of amorphous silicon to implement a photoconductive array and building up from such an array two neural network configurations.