Research And Application Of Extreme Learning Machine Algorithm For Classification

In recent years,artificial intelligence has been rapidly developed,machine learning as the core content of artificial intelligence has been widely used in various fields of artificial intelligence,such as pattern recognition,natural language processing,computer vision,expert systems,intelligent robots fields.Neural network as an important method of machine learning,researchers carried out in-depth study of neural networks.A variety of algorithms is presented based on the neural network and have achieved important applications in the regression problem and classification problem.As development of traditional single hidden layer feedforward neural network(SLFNs),extreme learning machine(ELM)is proposed by Huang et al.The ELM algorithm is characterized by randomly generating the input weights and the hidden layer node offset values.The solution of the traditional neural network is transformed from a nonlinear model to a linear model.The output weight is obtained by the least squares solution.And the traditional neural network solves the problem that obtaining the output weight is slow as it using the gradient descent.At the same time,researchers found that ELM has a higher speed of solution compared with traditional neural networks and better generalization ability.In recent years,researchers have applied extreme learning machines to classification,regression,feature extraction,clustering and other issues and achieved good results.However,extreme learning machine is still to be studied when facing the problems such as the irregular distribution of data,data with noise and outliers etc.,which can seriously affect the classification accuracy of extreme learning machine algorithm.In supervised learning,due to the limited data samples,ELM is not sufficient for generalization.In this paper,we present the research for ELM mainly on the two issues.The main research results are as follows:1)Based on Considering the influence of noise data,outlier data and the irregular distribution of data with the algorithm of extreme learning machine,and analyzing the activation functions in the extreme learning machine,proposes a robust extreme learning machine algorithm.2)For supervised learning,with lack of labeled data samples,proposes a regularized extreme learning machine based on manifold learning(LPELM).LPELM algorithm combines the discriminant information and the geometry of the data samples into the ELM model.Minimization of the internal divergence matrix is used to optimize the output weight of extreme learning machine and enhance the generalization ability of extreme learning machine.