Much of the power in neural-network processing lies in a network's ability to adapt to a given problem. The adaptation is accomplished by modifying its internal structure through some learning procedure. Neural network models may be classified in one of two types: The learning may be supervised by someone or something that indicates to the network what is expected of it or the network may be governed by a self-organizing process in which it automatically develops an internal state that in a self-consistent way reflects the properties of its input environment. Self-organizing systems need no a priori knowledge supplied by a supervisor and are particularly valuable when the task of the system depends only on some property of the input data itself. Here we describe a self-organizing optical system that extracts features from a collection of input patterns. The features are extracted according to a similarity criterion, which in our case is defined by an inner product. For example, if the collection of patterns consists merely of two images with orthogonal electric-field patterns, the system will recognize the orthogonality and the extracted features will be the images themselves. If instead the two pictures are not orthogonal but not very similar either, the system will find a pair of best features, which, in different linear combinations, make up the images. By contrast, if the two images are very similar, the system will decide that there is only one feature present and will indicate how much of that feature is present in each of the images.
© 1991 Optical Society of AmericaPDF Article