Fixed classifier pattern recognition using iteratively produced preprocessing

Pattern recognizers are often composed of two parts, the feature extractor and the classifier. This paper is a description of a pattern recognizer whereby the classifier learns first, and is then fixed, followed by learning by a preprocessor, which must learn how to predistort the input to the fixed classifier for proper recognition of the learning set. Learning the distortion is an iterative process whereby each vector of the training set must be examined for each iteration. Each iteration fixes the parameters for several fundamental distortions, and the use of a subset of all the distortions over all iterations constitutes a net distortion.