Search

This paper presents methods for training pattern (prototype) selection, class-specific feature selection and classification for automated learning. For training pattern selection, we propose a method of sampling that extracts a small number of representative training patterns (prototypes) from the dataset. The idea is to extract a set of prototype training patterns that represents each class region in a classification problem. In class-specific feature selection, we try to find a separate feature set for each class such that it is the best one to separate that class from the other classes. We then build a separate classifier for that class based on its own feature set. The paper also presents a new hypersphere classification algorithm. Hypersphere nets are similar to radial basis function (RBF) nets and belong to the group of kernel function nets. Polynomial time complexity of the methods is proven. Polynomial time complexity of learning algorithms is important to the field of neural networks. Computational results are provided for a number of well-known datasets. None of the parameters of the algorithm were fine tuned for any of the problems solved and this supports the idea of automation of learning methods. Automation of learning is crucial to wider deployment of learning technologies.

This paper presents a polynomial time algorithm for the construction and training of a class of multilayer perceptrons for classification. It uses linear programming models to incrementally generate the hidden layer in a restricted higher-order perceptron. Polynomial time complexity of the method is proven. Computational results are provided for several well-known applications in the areas of speech recognition, medical diagnosis, and target detection. In all cases, very small nets were created that had error rates similar to those reported so far.