Optimal structure analysis of universal learning network with multi-branches

Owing to the special characteristics of neural networks, the optimization of their structure is paid great attention. As the theoretical study and case analysis develop, the insufficiencies of networks with a single branch between nodes appear. (1) The loose structure when a single connection is used between nodes limits the application of neural networks in practice. Therefore, a compact network structure with multi-branches is required. (2) When a link branch from one node to another is eliminated, the signal transmission will be cut off completely between nodes. (3) Adaptability to the sharp change of inputs is poor. The paper, aiming at overcoming these insufficiencies, presents a method to create an optimal network structure. The basic idea of the proposed procedure is to introduce a Universal Learning Network (ULN) with multi-branches between nodes, then, meeting the demand of optimal structure, to eliminate unnecessary branches but not all of them. As a result, while the signal transmission is kept up, the network structure becomes compact. The problem of optimizing the structure of neural networks is to get the best balance between training precisely and generalization ability. Ensuring training precisely, the generalization ability of networks can be improved by the proposed procedure