Task decomposition and modular single-hidden-layer perceptron classifiers for multi-class learning problems

摘要: One of keys for multilayer perceptrons (MLPs) to solve the multi-class learning problems is how make them get good convergence and generalization performances merely through small-scale subsets, i.e., a small part original larger-scale data sets. This paper first decomposes an n-class problem into n two-class problems, then uses class-modular MLPs one by one. A MLP responsible forming decision boundaries its represented class, thus can be trained only samples from class some neighboring ones. When solving problem, has face with such unfavorable situations as unbalanced training data, locally sparse weak distribution regions, open boundaries. solutions that minority classes or in thin regions are virtually reinforced suitable enlargement factors. And next, effective range localized correction coefficient related class. In brief, this focuses on formation economic virtual balance imbalanced sets, localization MLPs. The results letter extended handwritten digital recognitions show proposed methods effective.