Recurrent Neural Networks with Multi-Branch Structure

Abstract

Universal Learning Networks (ULNs) provide a generalized framework to many kinds of structures of neural networks with supervised learning. Multi-Branch Neural Networks (MBNNs) which use the framework of ULNs have been already shown that they have better representation ability in feedforward neural networks (FNNs). Multi-Branch structure of MBNNs can be easily extended to recurrent neural networks (RNNs) because the characteristics of ULNs include the connection of multiple branches with arbitrary time delays. In this paper, therefore, RNNs with Multi-Branch structure are proposed and they show that their representation ability is better than conventional RNNs. RNNs can represent dynamical systems and are useful for time series prediction. The performance evaluation of RNNs with Multi-Branch structure was carried out using a benchmark of time series prediction. Simulation results showed that RNNs with Multi-Branch structure could obtain better performance than conventional RNNs, and also showed that they could improve the representation ability even if they are smaller sized networks.