Improvement of bidirectional recurrent neural network for learning long-term dependencies

Abstract

Bidirectional recurrent neural network (BRNN) is a non-causal generalization of recurrent neural networks (RNNs). Due to the problem of vanishing gradients, BRNN cannot learn long-term dependencies efficiently with gradient descent. To tackle the long-term dependency problem, we propose segmented-memory recurrent neural network (SM-RNN) and develop a bidirectional segmented-memory recurrent neural network(BSMRNN). We test the performance of BSMRNN on the problem of information latching. Our experimental results show that BSMRNN outperforms BRNN on long-term dependency problems.