Self-organizing broad network using information evaluation method

Broad network (BN) is commonly used as an efficient and computationally inexpensive machine learning method. The flat structure of BN ensures the efficiency of model learning, but in the mean time the structural characteristics also affect the network performances, such as generalization. However, it is challenging to improve the BN's generalization performance via adjusting its structure based on certain specific criteria. In this paper, to get the required network structure, a self-organizing BN based on the information evaluation method (IE-SOBN) is designed. First, a capability evaluation metric (CEM) is introduced into BN, which can evaluate the ability of hidden neurons to express and transmit data information. Then, the evaluation metric obtained by CEM is used as the basis for structure adjustment. Second, a structure adjustment mechanism (SAM) based on CEM is proposed to dynamically optimize the BN structure, leading to an improved generalization performance. Third, the convergence of the suggested IE-SOBN is theoretically analyzed to ensure its effectiveness. Finally, the advantages of IE-SOBN are verified by three real applications. The generalization performance and test accuracy of the proposed IE-SOBN are satisfactory in the experiments.