Computer generated planar holograms for optical neural network implementations
Paul E. Keller, Arthur F. Gmitro · Optical Society of America Annual Meeting · 1992
While there are several ways to implement the interconnections in an optical neural network, we have concentrated on the use of multifaceted planar holograms. Each facet or subhologram produces the synaptic connections from a single neuron output (optical emitter or modulator) to an array of neuron inputs (detectors). The hologram must accurately encode synaptic weights, have a high diffraction efficiency, utilize the least space–bandwidth product, and be fabricated by using standard electron-beam lithography and ion-etch technology. The various algorithms used to generate the interconnection hologram are discussed, and the algorithms are compared by their performance. The algorithms we have investigated for hologram generation include the cell-based techniques, error diffusion, the Gergchberg–Saxton process, simulated annealing, random-search error minimization, a hybrid of the Gergchberg–Saxton and random-search error minimization, and the genetic algorithm. The hybrid Gergchberg–Saxton/error minimization process has produced the highest interconnect accuracy and highest diffraction efficiency of the algorithms tested to date.