| As one of the most representative feedforward neural networks,the extreme learning machine has the advantages of less network parameters,no iteration and fast learning speed.However,because of its single hidden layer network structure is so simple that the learning ability of the extreme learning machine is limited,then the kernel extreme learning machine and the hierarchical extreme learning machine come out.But both of them have the disadvantages of being sensitive to noise,easy to overfit,and large amount of calculation.For these problems,this paper proposes two improved extreme learning machine networks based on the maximum correntropy criterion.Their performance is verified on the UCI data sets,and the applications of two improved extreme learning machine networks are studied.The specific research work is as follows:Firstly,the kernel extreme learning machine should have improved.Aiming at the problem that the extreme learning machine is sensitive to noise,the maximum correntropy criterion is introduced in the extreme learning machine and insteads of the minimum mean square error criterion to construct the objective function.In order to solve the problem that the single-kernel extreme learning machine is difficult to deal with the problem of heterogeneous or irregular datas,and at the same time improve the problem that the introduction of multi-kernel causes an increase in the amount of calculation,proposing the random scaling method to achieve multi-scale kernel.And a multi-scale kernel extreme learning machine based on the maximum correntropy criterion is derived,then its superiority in learning efficiency is verified through UCI data sets,and its practicality is verified in the prediction of the free calcium oxide content of cement clinker.Secondly,the hierarchical extreme learning machine should have improved.Inspired by improvements to the kernel extreme learning machine,in response to the problem that hierarchical extreme learning machine is prone to overfitting,the maximum correntropy criterion is introduced in the decision layer of the hierarchical extreme learning machine.In addition,applying kernel theory to the solution of the network output weight matrix in order to reducing the amount of calculation and improving the network’s nonlinear mapping ability.Then a kernel hierarchical extreme learning machine based on maximum correntropy criterion is constructed.Further,aiming at the problem of huge memory requirements when the hierarchical extreme learning machine is running,a learning method of batch coded is proposed,and the batch coded kernel hierarchical extreme learning machine under maximum correntropy criterion is proposed.Its network parameters are determined through simulation experiments,and comparison with hierarchical extreme learning machine verifies that it can effectively reduce memory requirements.Finally,the batch-coded kernel hierarchica extreme learning machine under the maximum correlation entropy criterion is applied to image classification.The data preprocessing process of the improved hierarchica extreme learning machine and the feature extraction principle of its unsupervised coding layer are elaborated.Then its decision layer as a classifier to batch learning on the image features obtained by the unsupervised coding layer on MNIST and NORB two image classification data sets,and compare with hierarchical extreme learning machine and other deep neural networks to verify its image classification performance. |
