Learning long-term dependencies in segmented memory recurrent neural networks

Abstract

Gradient descent learning algorithms for recurrent neural networks (RNNs) perform poorly on long-term dependency problems. In this paper, we propose a novel architecture called Segmented-Memory Recurrent Neural Network (SMRNN). The SMRNN is trained using an extended real time recurrent learning algorithm, which is gradient-based. We tested the SMRNN on the standard problem of information latching. Our implementation results indicate that gradient descent learning is more effective in SMRNN than in standard RNNs. © Springer-Verlag 2004.