| As we all know,the learning efficiency and learning speed of traditional neural networks are far below the requirements,which has become a major bottleneck for many applications.And the Broad Learning System is a recently proposed neural network model.Different from the traditional deep neural network model,the Broad Learning System has a flat network structure.Therefore,it can be trained quickly and efficiently.However,the model does not have the ability to select high-quality network nodes,and the network nodes added each time are randomly generated,so the model cannot effectively optimize the new hidden layer nodes by reducing the current network residuals.In view of the above two problems,this paper improves and expands the Broad Learning System.On the one hand,it analyzes the Broad Learning System from the perspective of information theory.By applying the newly proposed matrix-based R(?)nyi informationentropy,the mutual information between the feature node and the output node is analyzed,and the results show that the contributions of different nodes are different from each other.Therefore,a feature selection algorithm for orthogonal Broad Learning System based on mutual information is proposed.We perform experiments on classification and regression datasets to prove the effectiveness of the proposed algorithm.On the other hand,although the improved method proposed above makes the model have the feature selection ability,the improved model still does not have the learning optimization ability,so the Broad Learning System is expanded by combining the idea of reducing the residual of the current model every time.By reducing the residuals of the current network model each time,new hidden layer network nodes are optimized in sequence,and the Gaussian process latent variable model is used to extract features from the original data.Therefore,a bidirectional broad learning algorithmbased on Gaussian process latent variables is proposed.We also conducted experiments on the regression data set to prove the effectiveness of the proposed algorithm.This paper optimizes and expands the feature selection and generation methods of the network nodes of the Broad Learning System,and introduces information theory and Gaussian process latent variable model,and proposes an extension model based on the Broad Learning System,which improves the regression and classification capabilities of the extended model on data. |
