Generalization Of Random Weights Network Based On Uncertainty

In recent years, with the advent of the era of big data, the size and the complexity of the data are exponentially increasing. It has become a hot research topic in the field of machine learning and has important applications that how to effectively mine the knowledge with big data. For example, credit card fraud detection, weather prediction, medical diagnosis and so on are all well known applications of massive data mining.The classification problem is one of the main problems in the study of machine learning.By investigating the factors related to the generalization ability of classifiers, the main work of this paper is to find the relations between the factors and the generalization of classifier.Specifically, extreme learning machine is used as the classifier, we analyzed the influence of the complexity of the classification problem and the uncertainty of outputs on the generalization ability of classifier. Furthermore, a model is proposed which is used to describe the relations among uncertainty of outputs, complexity of the classification problems and the generalization ability of classifier. The following conclusions are verified by experiments:when the classification problem has high complexity, the generalization ability of classifier will increase with the increasing of the uncertainty of outputs; whereas, when the classification problem has low complexity, the generalization ability of classifier will reduce with the increasing of the uncertainty of outputs. The discoveries can offer help for improvement and evaluation of the generalization of the classifier.