Use of a Sparse Structure to Improve Learning Performance of Recurrent Neural Networks

Abstract

The objective of our study is to find out how a sparse structure affects the performance of a recurrent neural network (RNN). Only a few existing studies have dealt with the sparse structure of RNN with learning like Back Propagation Through Time (BPTT). In this paper, we propose a RNN with sparse connection and BPTT called Multiple time scale RNN (MTRNN). Then, we investigated how sparse connection affects generalization performance and noise robustness. In the experiments using data composed of alphabetic sequences, the MTRNN showed the best generalization performance when the connection rate was 40%. We also measured sparseness of neural activity and found out that sparseness of neural activity corresponds to generalization performance. These results means that sparse connection improved learning performance and sparseness of neural activity would be used as metrics of generalization performance.