The Implementations of High Robustness Recurrent Neural Networks

• Main Authors: Wen-Chung Hsu, 許文仲
• Other Authors: Hsien-Ju Ko
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/3gt6h8

碩士 === 亞洲大學 === 光電與通訊學系碩士在職專班 === 106 === In this thesis, a recurrent neural network based on high robustness state-space structure is proposed. The proposed RNN structure is local feedback. By using L2 sensitivity minimization, an optimal state-space realization can be synthesized with respect to finite precision implementations. The proposed structure is not only with minimal L2 sensitivity measure but also is L2-scaling constraint. This property may reduce the probability of overflow. Based upon back propagation learning algorithm, the proposed approach may lead to better performances when the RNN is implemented under finite precision devices. Finally, numerical examples are performed to illustrate the effectiveness of the proposed approach.