On Generalization Of Multi-Layer Feedforward Neural Network And Its Applications

The neural network has been studied for about many years and won successes in many fields, including pattern recognition, data mining and so on. Unfortunately, many problems, such as local minimum, overtraining, generalization and so on, are encountered in the theories, design and applications. The neural network generalization has been thought one of the most important targets, and attracted many researchers. In this paper, theories and methods about neural network generalization are discussed, and an iterative pruning algorithm is developed to improve the generalization. The computer simulation experiments on simulated data and multi-spectral images show the power ability of the algorithm.At first, the history of NN research is briefly reviewed, and the structure, function and BP algorithm are particularly introduced. Secondly, the definition and basic theories of NN generalization are presented. At the same time, neural network generalization ability and structural optimization in such respects as generalization theory and the prevailing methods for improving neural network generalization are summarized. Furthermore, the relationship of number of hidden neurons and.NN generalization is analyzed. And a new iterative pruning algorithm for improving the MLFNN generalization is developed, based on the idea of iteratively eliminating units and adjusting the remaining weights in such way that the generalization can be improved. In the end, the characteristics of high dimension multi-spectral data classification are briefly studied and MLFNN is used to classify the multi-spectral images. The new pruning algorithm is used to optimize the network structure and obtains high classification accuracy and good generalization. The experimental results obtained over simulated data and multi-spectral data demonstrate the effectiveness of the proposed algorithm.