| Broad learning system is a novel neural network framework,it relies on the broadwise expansion of feature nodes and enhancement nodes to achieve performance gains.Due to its computational properties of fast solving,functional properties of incremental learning,and structural properties of broadwise expansion,broad learning systems are attracting more and more attention from researchers.Currently,it has been used in a wide range of fields such as intelligent manufacturing,medical data analysis,communication engineering,and natural language processing.The success of the broad learning system is inextricably interrelated with the trendy demand for computing speed in big data.However,with further expansion of data size and complexity,the limited modeling capabilities of the broad learning system will not match the demands of future diverse task scenarios.This paper mainly focuses on key methods and techniques for enhancing broad learning systems and their variants in terms of expressiveness,generalization,and interpretability.On the current research process of the broad learning system,how to break the accuracy bottleneck of the model and enhance its expressiveness and generalization ability is the key to further improving the effectiveness of the model.Due to the scalability of the broad learning system,an important variant of this model,the fuzzy broad learning system,suffers from similar problems.In addition to the above issues,since fuzzy broad learning systems introduce fuzzy subsystems,how to trade off the accuracy and inference capability becomes another important goal in the current development of fuzzy broad learning systems.In summary,the research questions and content of this paper can be summarized as follows:(1)Aiming at the problem that the contribution of the enhancement layer of the fuzzy broad learning system is unknown and the combination of parameters is sub-optimal,a fuzzy broad learning system based on sparse enhancement and multiple clustering is proposed.Multiple fuzzy subsystems are embedded in a broad learning system,and different of them make work coordinately with each other for improving performance.However,the stochastic mapping in the enhancement layer of this model does not produce a clear and significant gain for the model as a whole.Therefore,this paper proposes a principal component-based sparse autoencoder for extracting the key features in the enhancement layer while reconstructing the effective information lost due to dimensionality reduction.In addition,this paper also determines the optimal performance under the current structure by exploring different components and parameters of the fuzzy broad learning system.The experimental results show that the different combined models obtained in this paper can further improve the accuracy of the fuzzy broad learning system.Among them,EPFCM-SEFBLS achieves91.43%,99.88%,87.84%,99.16% and 95.18% accuracy on Heart,Wine,Vehicle,WDBC and IS datasets,respectively,which validates the effectiveness of the proposed model.(2)Aiming at the problem of the difficult trade-off between accuracy and inference ability of fuzzy broad learning systems,a more compact mixed fuzzy width learning system is proposed.As a novel neuro-fuzzy model,the fuzzy broad learning system has more fuzzy systems compared to other neuro-fuzzy models,which significantly increases the number of fuzzy rules,leading to the model under the dilemma of rule explosion and rule duplication.This paper introduces key methods from deep learning into a fuzzy broad learning system.Dropout techniques are used to reduce the number of fuzzy rules and fuzzy subsystems.An attention mechanism based on a sparse autoencoder is designed,which enhances the enhancement layer’s effectiveness and compensates for the performance degradation of the fuzzy subsystem due to rule clipping.Finally,an expert mixture model is used to facilitate the integration and coordination of different techniques and methods.The experimental results show that the model can better trade-off accuracy and inference while improving generalization capabilities.(3)Aiming at the problem of limited modelling capability of the broad learning system,a cascaded neural network framework with sparse polynomial weights-based RBF neural networks and attention broad learning system is proposed.When faced with complex and entangled data,it is a struggle to achieve satisfactory performance results.This paper enhances the local mapping capability and feature layer representation of the width learning system based on sparse polynomial RBF neural networks and attention mechanism-related methods.A two-stage training method with unsupervised training and supervised training is used to improve the generalization ability of the model in a stepwise learning manner and to produce regularization effects.The experiments show that the proposed neural network framework is a broad learning model with better mapping capability,which can improve the model performance while having more significant noise-resistant robustness.Experimenting on the Yale database with 70% salt and pepper noise,the proposed model can still achieve70.67% recognition accuracy,which proves the effectiveness of the model against noise. |
