Incremental Extreme Learning Machine Based On Error Constraint

In big data era,various artificial intelligence technologies have been developed rapidly.To improve machine learning performance,it is the core technology of artificial intelligence to approximate the learning behavior of human brain.Deep learning promotes the development of the artificial neural network,and also stimulates the research of extreme learning machine,a new type of learning method.Therefore,we will carry out in-depth research and discussion on the extreme machine learning(ELM).In this thesis,how to build the optimal network structure to improve learning performance and generalization ability based on ELM is studied and it can explore the stability of ELM under the condition of data changes.The main works are as follows.(1)Firstly,a general overview of machine learning and extreme learning machine,including its development background and significance,related theories,the main research status,at the same time,and the challenge of extreme learning machine is also presented.(2)To select the optimal number of hidden layer nodes,we deeply study an extreme learning machine with adaptive growth of hidden nodes and incremental updating of output weights(AIE-ELM).And the theoretical basis of adaptive incremental learning machine is then introduced in details,as well as the framework of learning mechanism and learning process.The results of comparative experiments show the effectiveness of AIE-ELM.The shortcoming of the AIE-ELM is its learning instability and we made a detailed analysis of this problem.(3)Finally,aiming at the problem of learning instability of the AIE-ELM,which is caused by changes in the data,we propose an improvement method based on local generalization error model(AIEL-ELM).By adding a constraint of local generalization error,we set up the upper limit of mean square error of the neighborhood of training data sample,so that we can reduce the output sensitivity to the input varying.The experimental results show that the improved extreme learning machine can improve classification accuracy and regression error,and the generalization ability is also enhanced.