Feedforward Sequential Memory Networks: A New Structure to Learn Long-term Dependency

In this paper, we propose a novel neural network structure, namely \emph{feedforward sequential memory networks (FSMN)}, to model long-term dependency in time series without using recurrent feedback. The proposed FSMN is a standard fully-connected feedforward neural network equipped with some learnable memory blocks in its hidden layers. The memory blocks use a tapped-delay line structure to encod

[KSKSKSKS2]

シーケンシャルなデータを扱うための Neural Network の高速化モデル。構造上確かにRNNよりは早く学習できそう