Structural optimisation algorithm of weight correlation analysis for DBN

Abstract

The deep belief network (DBN) algorithm has slow convergence rate and low learning efficiency, and it is sensitive to model hyper-parameters. In view of the fact that DBN is a learning network model with a large number of network nodes and multiple hidden layers, this paper proposes a weight correlation analysis algorithm for DBN (WCA-DBN) to determine the approximately optimal network structure of DBN. This algorithm establishes a PS orthogonal projection space based on the 'arbitrary node set' using the linear discriminant analysis (LDA) method, projects the 'balanced node set' onto the PS plane, and solves the weights of the network nodes between sets through the correlation principle, to adjust the network structure of the hidden layer adaptively. The real-time traffic flow of six different types of roads in Jinan City, Shandong Province was selected as experimental materials to verify that the WCA-DBN algorithm has more obvious advantages over traditional algorithms in terms of accuracy and time complexity.