Ensemble System of Multiple Recurrent Neural Networks

摘要：

Recurrent neural networks made up of long short term memory blocks is a kind of newly appeared network for offline character ecognition. Compared with the hidden markov models, which are currently widely used, RNNs dont have the limitations that HMMs have, and have achieved better experiment results.This paper improves the CTC output layer of the RNN by making it output the TOPN results list with an acceptable time complexity. Then a two-level combination scheme is used to combine multiple RNNs. Experiments results show that the average recognition rate of the first level combination is higher by 27.31％ than the member RNNs. The second-level combination is proved to be efficient in reducing the standard deviation of the recognition rates.

关键词： Recurrent Neural Networks Long Short Term Memory Multiple Classifiers Ensemble Offiine Character Recognition

作者: Zhang Liang Huang Shuguang Shi Zhaoxiang Tang Heping Hu Ronggui

作者单位: Hefei Electronic Engineering Institute Hefei, Anhui, China

会议类型: 国际会议

会议名称: The 10th International Conference on Intelligent Technologies(第十届智慧科技国际会议 InTech09)

会议地点: 桂林

会议语种:英文

页码: 647-651

在线出版日期: 2009-12-12（万方平台首次上网日期，不代表论文的发表时间）