Design of a locally interconnected modular neural network

Abstract

In order to design a model that is similar to the brain-like neural network while simultaneously enhance the learning performance of neural network, we propose a locally interconnected modular neural network (LI-MNN) model by adding inhibitory connections between the hidden nodes of different subnetworks. Firstly, to be more approximate to the pattern of brain functional division, the LI-MNN applies the Entropy Weighting k-means algorithm to partition all the features into several feature subsets. Secondly, each subset uses a BP neural network to learn and the excitatory hidden nodes in different subnetwork are connected with each other. The parameters in all subnetworks are learned by the improved Levenberg-Marquardt algorithm. Finally, the proposed model is tested on several benchmark datasets and a practical dataset in sewage treatment process, and the experimental results show that LI-MNN possesses superior generalization ability when compared with other MNN models.