Research On Regularized Extreme Learning Machine Based On Genetic Algorithm And Convolution Neural Network

As a special single hidden layer feedforward neural network,the extreme learning machine randomly gives the weights of the input layer,and computes the hidden layer weights by the least square method in onetime.It has fast learning speed and good approximation ability.On the basis of ELM,regularization theory is introduced into the regularized extreme learning machine(RELM),In the RELM,the empirical risk and structural risk in the statistical theory are all considered,and this two risks are balanced by the risk ratio parameter.The generalization performance of RELM is better than that of ELM.RELM is widely applied to classification,regression and prediction,and it achieves great results.Although RELM enhances its generalization performance and the stability of the original ELM by introducing the regularization theory,the RELM acquires the risk ratio parameters by the methods,such as random acquisition and cross validation,and it does not guarantee to get the optimal risk ratio.At the same time,similar to ELM,RELM adopts random network parameters and a free iterative method learning for training classification,which increases the uncertainty of the whole network,and can not adjust the network according to the actual situations.In view of these problems,this paper intends to optimize the RELM in order to get better performance of RELM and apply it to the practical field.The main research work of this paper is as follows.(1)Firstly,a regularized extreme learning machine based on genetic algorithm(GA-RELM)is proposed in this paper.We combine the genetic algorithm with the regularized extreme learning machine and we use the genetic algorithm to adaptively acquire the proportional parameters of the two risks:structural risk and experience risk.At the same time,we apply the proposed GA-RELM to the face recognition,and show the simulation experiments in Matlab R2014a.And the proposed GA-RELM is compared with the other methods.Simulation results and comparative analysis show that GA-RELM has good performance.(2)Secondly,we combine the convolution neural network(CNN)with RELM and propose a regularized extreme learning machine model based on convolution neural network(CNN-RELM).We use the CNN-RELM as classifier in face recognition.In this model,the convolution neural network is trained first,and the parameters of the CNN model are fixed after the learning target accuracy is reached,and the full connection layer of the convolution neural network is replaced by the regularized extreme learning machine,and our CNN-RELM model is obtained.We have simulated the face data set by the proposed CNN-RELM model and compared with the other methods.The simulation results show that the recognition rate of the proposed method is high.In this paper,two improved models of RELM are proposed and applied to the face recognition.The experiments show that the two RELM models can effectively improve the performance of the original RELM,and this has some theoretical and practical significance.