Optical learning machines

While learning has proven useful for some classes of problems in sequential machines, many researchers believe that the real power in learning cannot be accomplished without also mimicking natures ability to provide massive interconnections for parallel operation. Such connectionist architectures or neural networks are difficult to construct in conventional electronic technologies. The uncharged nature of photons relative to electrons leads to superior performance for moving high bandwidth signals in massive numbers in close proximity. This paper considers systems that perform learning optically. Therefore, we restrict ourselves to fairly simple algorithms with analog optical systems. Both supervised and unsupervised learning systems are reviewed. Supervised systems, because they use a teacher, are often subdivided into learning and recall. Some researchers use optics for recall and electronic computers for learning. We restrict ourselves to optical learning because it is often the case that learning in electronics is too slow and the system loses adaptability if it has to stop for long periods to retrain.