Designing a Pattern Recognition Neural Network with a Reject Output and Many Sets of Weights and Biases

Abstract

Most neural networks that have been designed to solve the problems of pattern recognition use a supervised training method with a training data set. This data set contains examples of input patterns together with the corresponding output results, and the neural network learns to infer the relationship between input patterns and output results through training. In supervised training, we often try to find out a set of weights and biases for the neural network in order to classify all patterns in the training data set. In general, training with a larger training data set can reduce the recognizing error rate. However, it would be difficult to find out a good design of neural network that will be able to learn all patterns in a large training data set, because it usually contains some patterns that are difficult to classify. Even if network layers and neurons were added more, there are still some misclassified patterns after a long time training process. The number of these patterns will increase when the size of the training data set is enlarged. If the neural network has to recognize a pattern that approximates in shape to one of the misclassified patterns, the recognition result will be incorrect. Furthermore, if a new pattern is updated, which approximates in shape to one of the misclassified patterns in the old training data set, the neural network may not still classify it, and it will become a new misclassified; thus, the error rate will increase. In this chapter, we introduce a new design of pattern recognition neural network that has a simple structure but is still able to classify almost all training patterns exactly. The neural network is designed with an especial output that is called “Reject output”. With this output, a large training data set can be separated into some parts, and with a smaller number of patterns in each part, they can be classified by the neural network more easily using a distinct set of weights and biases. Additionally, we also design a training method with some phases, which helps the neural network with the reject output to find out not only one but many sets of weights and biases for classifying almost all the training patterns. All the sets of weights and biases have to be kept in the order that they have been received from the training process. Moreover, the reject output is also used to control the updating process for new patterns more easily. With the reject output, the pattern recognition neural network can produce not only correct or incorrect results but also reject results; therefore, it can control the recognizing rejection and reduce the error rate. Open Access Database www.i-techonline.com