Nonlinear dynamics in optics and neural networks
Dana Z. Anderson, Claus Benkert, Don Montgomery, Mark E. Saffman · Optical Society of America Annual Meeting · 1991
By tailoring the interaction among modes of a photorefractive ring oscillator we have been able to produce a variety of neural network dynamics including winner-take-all and "voters paradox" behavior. Our systems take advantage of the mathematical similarity between neural network models and mode competition in lasers. In the present case, mode competition is mediated by two-beam coupling in photorefractive materials. Here we consider how tailored mode interactions can also give rise to self-organized information processing. We demonstrate a system that learns to demultiplex two signals from a multimode fiber, and another system that extracts features from a set of images, then subsequently performs feature correlation on a test image. Both systems perform their task by recognizing that the input environment consists of spatially and temporally independent components. There is very little a priori information given to the systems; instead, they self-organize based on the differences among the signals or images. Although the demonstrated systems are quite specific, the self-organizing principles demonstrated are quite general and may be used for a variety of signal processing and optical sensor tasks.