Abstract

An optical implementation of a neural net computer consists of two basic components: neurons and connections. The neurons are simple nonlinear processing elements (e.g. thresholding units) that accept inputs from other neurons and produce a single output that is broadcast to many other neurons. Typically we think of each neuron being connected to thousands others. Hence the number of connections in a network is much larger than the number of neurons. This simple fact is the principal motivation for considering optical implementations of neural nets [1]. The basic approach we have adopted is shown in Fig. 1. The neurons are arranged in a planar configuration and interconnected with optical elements (holograms or masks). Several emerging optical technologies can be considered for the implementation of the two basic components. We have come to the conclusion that the most promising technology for the implementation of neurons is optoelectronics; a two dimensional array of LEDs, a detector adjacent to each LED and a saturating amplifier connecting the two [2,3]. We are considering two possiblities for performing the connections: optical memory disks and volume holograms [2]. In this paper we examine the advantages of using volume holograms as opposed to planar media for storing the interconnect pattern, we present methods for achieving different types of arbitary global interconnections, and we present experimental results using a photerefractive crystal (LiNbO₃) to implement modifiable synapses.

© 1987 Optical Society of America