Abstract

While volume holograms offer a powerful means for providing interconnections for many information processing tasks, the adaptable interconnection capability necessary for implementing learning algorithms for neural networks finds an almost ideal realization in photorefractive crystals. The dynamic holograms in which gratings are built up in real time without the need for separate development cycles make them appealing for neural implementations.¹-³ To implement learning algorithms such as the Perceptron or the Adaline, a means for making additive as well as subtractive changes to gratings selectively is necessary and crucial to ensure convergent behavior. Recently³ we described an architecture in which the Stokes principle of reversibility for light was used in a simple optical system to realize the subtractive as well as additive holographic modifications. Selective subtraction of grating amplitudes was performed using what is basically a double Mach-Zehnder interferometer.⁴ A coherent laser beam is split into two paths by a beam splitter. One path serves as the reference beam while the other is expanded to pass through the spatial light modulator containing an image. The two beams intersect within a photorefractive crystal to write a holographic grating. By supplying another laser beam to the beam splitter and arranging its two resultant beams to traverse exactly the same paths as the first beam, a grating identical to that written already but out of phase by 180° can be written. Thus, simply by controlling which beam is used to expose the hologram one can either add to or subtract from the holograms present in the crystal.

© 1989 Optical Society of America