Noise Tolerating Capability Of Bagging-based Neural Networks And Their Application

It is known that an important factor for the fidelity of a neural network is the quality of the training data.Although neural networks are known to have certain ability to learn from noisy data,performance degradation is very hard to avoid.It is often difficult and sometimes impossible to completely exclude noise,which could be introduced in collecting tasks such as selecting,measuring and recording.Therefore,it is necessary to explore a way for improving the quality of training samples.In this paper,based on the features of the bagging algorithm and neural network ensembles,we investigate the noise tolerating capability of bagging-based neural network ensembles.When the noise level in a data set is reasonably small(＜35%),the architecture, weight settings and output values of the network are determined mainly by the non-noisy data. This provides a mechanism to correct noisy data by bagging-based neural network ensembles. Numerical experiments show that bagging-based neural network ensemble has high noise tolerating and filtering capability.We also use the method to solve a mathematical formula recognition problem.The contents of the paper are as follows:Chapter 1 reviews the history and basic knowledge of neural networks,neural network ensembles and three typical methods of using neural network to deal with different types of noisy data.Chapter 2 focuses on the study of the noise tolerating and filtering capability of bagging-based neural network ensembles for small training sample sets.Numerical experiments show that when the training samples contain noise ratio is less than 35%,the method is effective for the vast majority of the sample collections.Chapter 3 discusses a recognizer for mathematical formula recognition based on neural network ensembles.Our experiments show that the recognition rate of the bagging-based neural network ensemble is about 6%higher than that of a single BP neural network.