Abstract

One approach to implementing the space-variant interconnections needed for optical neural networks is to use a multifaceted hologram where each node has a dedicated subhologram, or facet, that encodes the interconnection pattern for that node. This approach was used to build an 8 × 8 node neural network for a simple associative memory problem and an 8 × 8 node neural network for a sixteen input winner-take-all network. The design approach for these holograms is presented, and the performance of the system in terms of diffraction efficiency and interconnection accuracy is discussed.

© 1988 Optical Society of America