Arabic Vowel Classification via QR Factorization

摘要

The generalization ability of a neural network is influenced by the size and efficiency of the training set. As a result, a key factor in any neural network application is the training set selection. In this paper, a new algorithm for optimizing this selection is proposed. The algorithm, based on the QR factorization of the training patterns, provides a systematic selection of an optimum training set. Consequently, the computational effort in the training phase of a neural network is significantly reduced and yet a satisfactory level of generalization is maintained. The potnetial of the algorithm is demonstrated on the problem of Arabic vowel classification.