Improve The Generalization Method And Application Of Extreme Learning Machine

The focus of traditional ELM optimization algorithms is on the output weight of the hidden layer,ignoring the distribution of the input and output of the hidden layer.In view of this it is easy to overlook direction,this paper presents Extreme Learning Machine model based on a Gaussian distribution affine parameter optimization,the specific contents are as follows:In order to solve the problem that the output of hidden layer is forced to obey the uniform distribution,the extreme learning machine for optimized affine transformation based on Gaussian distribution(GAT-ELM)model is proposed.In the model,the affine parameters are introduced,and the gradient descent algorithm is used to optimize the data input to Gaussian distribution.The output data of the hidden layer obtained by sigmoid activation function mapping also obey the Gaussian distribution.The model adopts the loo(leave one out)cross validation strategy in the RELM(Regulated ELM)algorithm,and SVD(Singular Value Decision)algorithm is introduced to calculate the output weight of the hidden layer.The model has good generalization performance and improves the robustness of the model.Six image data sets and 12 UCI data sets are used to divide them into small,medium and large sample sizes.The comprehensive and systematic experiments show that the proposed model has good generalization ability and stability for practical application.Aiming at the problem that general ELM models with large sample size and high complexity data sets have poor generalization ability,a Multiple Kernel Extreme Learning Machine With Affine Transformation(ATMK-ELM)model with affine parameters is proposed.In the model,affine parameters are introduced on the basis of the Multiple Kernel extreme learning machine(MK-ELM),and the gradient descent algorithm is used for optimization,so that the sample data obeys the Gaussian distribution and eliminates outliers.The model uses the linear mapping method of the hidden layer input data in the AT-ELM(Affine Transformation ELM)algorithm to make the original data obey the Gaussian distribution,and perform multi-core learning on the optimized data,without the need to explicitly define the feature space and mapping function In more complex data sets such as classification and regression,kernel mapping is used instead of random mapping to improve the generalization performance and poor stability caused by random assignment of hidden layer nodes.The experimental results show that in complex data with large sample size Concentrated,the ATMK-ELM algorithm performs better than ELM,RELM,MK-ELM and AT-ELM.