Research On Evolutionary Broad Learning System Based On Broad Auto-encoder

At present,machine learning technology has been widely used in real life and brought great convenience to human life.Among them Deep learning technology is the most widely used,but deep learning algorithms are often restricted to training resource limitations due to too many parameters that need to be trained,and because it uses back-propagation iterative algorithms to optimize parameters,model training often costs a lot of time.The newly proposed Broad Learn-ing System(BLS)provides an alternative method of deep learning which saves time-consuming training process and powerful computing resources.However,the randomly generated feature nodes and a large number of enhancement nodes in BLS network may reduce the performance of the final classifier.This article focuses on the problem of Broad Learning System(BLS)in ex-tracting data features and proposes two different types of algorithm models from the perspective of supervised and unsupervised feature extraction,respectively.First,from the perspective of supervised feature extraction and aiming at the unpredictable problems in randomly generated feature nodes in BLS,an adaptive feature node evolution algo-rithm(AFNE)is proposed to extract better features.Based on AFNE,a Broad Learning System based on adaptive feature evolution(AD-BLS)is also proposed.In addition,BLS network usu-ally requires a large number of enhancement node parameters to ensure its performance which can easily lead to redundant dependencies between features and the final model degradation.Therefore,this paper also proposes a Pruned Broad Learning System based on adaptive feature evolution(ADP-BLS)to prune the network enhancement layer,delete redundant nodes,reduce network scale,and increase generalization capabilities.The proposed method ADP-BLS can improve the accuracy and generalization ability of the final classifier through the evolution of feature nodes and pruning of enhancement nodes.In order to verify the performance of the ADP-BLS network,we carried out a lot of experiments to prove it.Meanwhile,module abla-tion experiments are conducted to verify the improved effect of the algorithms added to the BLS network.Secondly,from the perspective of unsupervised feature extraction and aiming at the time-consuming defects of traditional deep autoencoder training based on backpropagation algorithm.This paper also proposes a new type of self-encoding network feature extractor based on Broad Learning System which includes the BLS self-encoding network(BLS-AE),the stacked BLS self-encoding network(ML-BLS)and the sparse BLS self-encoding network(H-BLS)by train-ing with L1 regularization method.These self-encoding networks built on the basis of BLS still retain the advantages of fast training in BLS network,and overcome the time-consuming short-comings of traditional self-encoding networks that require iterative optimization after a large number of parameters are initialized.At the same time,the higher-level abstract features of the input data can be learned through the progressive encoding and decoding process,and using L1 regularization to train the parameters can further enhance the robustness of the encoded features.Finally,from the perspectives of unsupervised feature extraction and supervised model classification,the main research work in this article are integrated as an end-to-end Evolutionary Broad Learning System from feature extraction to classification based on BLS Autoencoders.After using ML-BLS and H-BLS as feature extractors to extract features,they are input to ADP-BLS network for the final classification task.A large number of application experiments on real-world data sets show that the algorithm proposed in this paper not only has the advantage of fast training,but also has good performance compared with the traditional mainstream algorithms.