Programmable multilayer optical neural networks with asymmetric interconnection weights

A hybrid optical-electronic architecture is described which uses gratings stored in a photorefractive crystal to interconnect neurons using coherent light beams. Neurons are represented by pixels on two liquid crystal light valves. Threshold functions are implemented electronically using lookup tables. Area transformation of two neuron planes are used to implement arbitrary multilayer neural networks with asymmetric interconnection weights using a single hologram to store the interconnection pattern. Some factors limiting storage capacity and processing speed are discussed.<>