一种基于前馈序列记忆神经网络的改进方法

An Improved Method Based on Feedforward Sequence Memory Neural Network

针对具有时序性的信号的分析和建模,主流的RNN、LSTM由于反馈连接的影响,在学习效率和稳定上有所不足。本文基于标准的前馈神经网络,借鉴滤波器中的抽头延迟线结构,提出一种改进的前馈序列记忆神经网络FSMN(cFSMN)和深层cFSMN(Deep-cFSMN),实现时序的音视频信号快速建模,减少了反馈连接,具有更高的学习速率和更好的稳定性。

For the analysis and modeling of sequential signals,the mainstream RNN and LSTM have some shortcomings in learning efficiency and stability due to the influence of feedback connection.Based on the standard feedforward neural network and the takeout delay line structure in the filter,this paper proposes an improved feedforward sequential memory neural network fsmn(cfsmn)and deep cfsmn(deep cfsmn)to achieve sequential sound.Video signal fast modeling reduces feedback connection,has higher learning speed and better stability.