The authors consider the optical implementation of learning networks using volume holographic interconnections in photorefractive crystals. The use of volume holograms permits the storage of a very large number of interconnections per unit volume, and the use of photorefractive crystals permits the dynamic modification of these connections, thus allowing the implementation of learning algorithms. The authors first briefly review the major types of learning algorithms that are being used in neural network models. They then estimate the maximum number of holographic gratings that can simultaneously exist in a photorefractive crystal. Since in an optical implementation each grating corresponds to a separate interconnection between two neurons, this estimate gives the density of connections that are achievable. They consider how the modulation depth of each grating (or equivalently the strength of each connection) can be controlled through the implementation of learning algorithms. Two related issues are investigated: the optical architectures that implement different learning algorithms and the reconciliation of physical mechanisms that are involved in the recording of holograms in photorefractive crystals with the dynamics of the learning procedures in neural networks.
LEARNING IN OPTICAL NEURAL COMPUTERS.
Abstract
Year