Improvement of Bidirectional Recurrent Neural Network for Learning Long-Term Dependencies

Bidirectional recurrent neural network(BRNN) is a noncausal generalization of recurrent neural network(RNN). It can not learn remote information efficiently due to the problem of vanishing gradients. To limit this problem, we propose segmented-memory recurrent neural network(SMRNN) and replace the standard RNNs in BRNN with SMRNNs. The resulting architecture is called bidirectional segmented-memory recurrent neural network(BSMRNN). We test the BSMRNN on the problem of information latching and provide results showing that BSMRNN performs better on long-term dependency problem than BRNN.

• Type: Conference
• Year: 2004
• Where: ICPR
• Authors: Jinmiao Chen, Narendra S. Chaudhari
• Keywords: Computer Vision, ICPR 2004, Recurrent Neural Network(BRNN, Recurrent Neural Network(RNN, Recurrent Neural Network(SMRNN