Abstract

Elementary holographic gratings can be used to implement interconnection links betweeen individual processing elements of two distinct planes in a multi-layered optical neural network. In such networks, one issue that has a direct impact on the development of learning machines is the capability of continuously modifying a given interconnection strength (or weight) without affecting the others, when the gratings share the same volume in the photorefractive crystal (i.e., frequency-multiplexed gratings). In the following, we extend the principle of coherent erasure by the double-exposure technique to the case of elementary gratings that implement real-time optical interconnections in photorefractive materials. The effect of continuously varying the phase shift between the two recorded gratings on the diffraction efficiency is quantified, and shown to be applicable to the simulation of synapses with programmable variable weights, as would be required in a learning neural network. Issues that relate to fan-in and fan-out capabilities, which ultimately determine the achievable level of parallelism and cascadability in such processing architectures, are also addressed. An experimental interconnection system based on two-dimensional liquid crystal phase and amplitude modulators and photorefractive recording is described. Finally, an extension of the double-exposure technique to time-average erasure is then discussed.

© 1989 Optical Society of America