29 November 2019 14:48 GRU is a popular variant of LSTM which replaces the forget gate and the input gate with only one update gate GRU achieves the similar performance in multiple tasks with less computation LSTM 通过门控机制使循环神经网络不仅能记忆过去的信息,同时还能选择性地忘记一些
LSTM的结构中每个时刻的隐层包含了多个memory blocks(一般我们采用一个block),每个block包含了包含一个Cell(有多个memory cell组成)和三个gate,一个基础的结构示例如下图: 一个memory cell只能产出一个标量值,一个block能产出一个向量. The nomenclature is a bit confusing, but goes back to the original LSTM paper: an LSTM "cell" is s