Research And Application Of Broad Learning Algorithm

In the practical modeling process,the quality of the model is mainly determined by the mapping relationship between the target value and the related features,while many complex systems have problems such as nonlinearity and strong coupling,which make it difficult to construct high-precision models.The proposed neural network provides an effective solution to the above problems.However,with the increase of the number of neural network layers,the parameters of the model also increase.Therefore,the optimization of parameters and the excessive depth of the network structure have become the main problems of current neural networks.For this problem,the broad learning system not only overcomes the lengthy parameter iteration process,but also can gradually increase the network nodes and update the nodes without re-modeling.And the algorithm has shorter training time and better generalization ability.In order to further improve the performance of the broad learning system,this paper has improved it,and verified the performance of the improved algorithm on image classification and data regression.The specific research results and practical significance are described below:(1)This paper proposes a local receptive fields based broad learning system(BLS-LRF),which combines the broad learning system(BLS)and local receptive fields(LRF).In order to test the effectiveness of the algorithm,it is applied to the classic MNIST and NORB image classification datasets.Through comparative experiments with classic algorithms such as convolutional neural network algorithms(CNN)and residual neural networks(ResNet),the simulation results show that BLS-LRF algorithm not only improves the generalization performance of the algorithm,but also greatly shortens the training and learning process.(2)This paper proposes a multi-parallel broad learning system(MPBLS),which should solve the problems of image classification and data regression.The simulation results show that in the problem of image classification,MPBLS improves the classification accuracy of the algorithm compared with other classical algorithms and reduces the time required for algorithm training at the same time.in the data regression problem,compared with other algorithms,the MPBLS not only shows good data fittingability and generalization ability,but also through the excellent performance of small data,it can be seen that the size dependence of MPBLS on training samples is small.