Holographic learning to classify optically preprocessed signals

Abstract

Linear optical processors (such as optical spectrum analyzers, correlators, optical Wigner processors, and ambiguity function processors)1 can rapidly extract classification features from wide bandwidth signals. However, with dimension-increasing processing (such as the ambiguity function) the output information rate can massively exceed the input capacity of digital computers used for classification. An optical classifier, such as an adaptive optical neural network,2 however, can potentially provide a throughput rate to match the output of the optical feature extractor. As a demonstration of this concept, a broadband communications signal classfier was constructed by cascading an acousto-optic spectrum analyzer with an adaptive holographic pattern classifier using a photorefractive crystal of Fe:LiNbOr. Experimental results obtained by using this two-stage processor asa shift invariant classifier are included. This configuration requires an error-driven learning pathway for weight modification to implement an adaptive classifier. A multilayer system to classify radar returns from an isolated aircraft by using an adaptive neural network classifying the output of an optical ambiguity function processor is proposed.