Research On Robust Broad Learning System And Its Application

In recent years,the development of machine learning and artificial intelligence has received extensive attention from academia.As one of the most attractive neural network models,deep learning has shown excellent performance in many fields.Although the complex multilayer network structure of deep learning exhibits excellent learning capability,it also poses some obstacles to model design and training.The network design of deep learning relies on experience and experiments,and model training consumes a lot of computing resources,which limits the further development of deep learning.The emergence of broad learning systems provides a new idea for the development of machine learning.Broad learning system adopts a novel shallow network structure,where the network parameters are greatly reduced compared to the deep learning model,and it does not require an iterative process for model training.This gives the broad learning system a great advantage in training speed.However,with further research,it is found that the performance of the broad learning system is severely degraded by noise and outlier interference.Its network uses the least squares based loss function that is sensitive to noise and outliers,making the data contaminated by noise and outliers severely disrupt the model training process of the broad learning system.To solve the above problems,this paper proposes two robust broad learning systems,which are summarized as follows:(1)For the case of training data contaminated with noise and outliers,this paper proposes a broad network called the self-paced broad learning system.The self-paced broad learning system incorporates the self-paced learning mechanism into its network,adaptively assigning appropriate priority weight to each training data.The weight assigned to the training data will guide the training of the self-paced broad learning system to rely more on samples with high confidence.The self-paced broad learning system achieves the identification of reliable data for model training in a complex noisy environment and effectively suppresses the negative effects arising from noise and outliers.In addition,this paper proposes the incremental learning algorithms corresponding to the self-paced broad learning system,which achieves the ability to improve the performance of the system in noisy environments by expanding the network structure or adding additional training data.Extensive experiments demonstrate the effectiveness of the proposed algorithm.(2)The robustness of the broad learning system is constrained by the least squares based loss function in the model.Accordingly,this paper further proposes a Cauchy regularized broad learning system.The Cauchy loss function replaces the least squares based loss function design in the model,which effectively improves the capability of the broad learning system against noise interference.The Cauchy regularized broad learning system also achieves a flexible model update capability,and the network can improve its performance by learning additional training data without retraining the whole model.Through experiments,the robustness of the proposed algorithm has been verified.