Multibranch structure and its localized property in layered neural networks

Neural networks (NNs) can solve only simple problems if the network size is too small, but increasing the network size is costly in terms of memory space and calculation time. Thus, we have studied how to construct a network structure with high performance and low cost in space and time. One solution is a multibranch structure. Conventional NNs use the single-branch structure for connections, while the multibranch structure has multiple branches between nodes. In this paper, a new method which enables the multibranch NNs to have the localized property is proposed. It is well known that RBF networks have the localized property, which makes it possible to approximate functions faster than sigmoidal NNs. By using the multibranch structure having the localized property of RBF networks, NNs can obtain superior performance while maintaining lower costs in space and time. Simulation results of function approximations and a classification problem are presented to illustrate the effectiveness of multibranch NNs.