Trainable optically programmed neural network
Richard G. Stearns · Applied Optics · 1992
The use of a two-dimensional photoconductive sensor array that is coupled with a liquid-crystal display is examined as a basis for the hardware implementation of neural network architectures. An array of 128 x 128 amorphous-silicon photoconductors of lateral geometry is fabricated and mated to a liquid-crystal device, the latter of which is used to optically program the conductances of the sensor array. Implementation of a three-layer perceptron network that uses the combined sensor and display devices is described. As an example of its operation, the system is trained to recognize images of numeric characters, with excellent results.