本文共 501 字，大约阅读时间需要 1 分钟。

Long Short-term memory is a kind of recurrent neural networks. It has additional gates and its' architecture is bellow:

This architecture can avoid the gradient vanishing and keep the long-term memory, so this is a very powerful neural network.

We can get these formulations:

The symbol   denotes the indexs of all units,  denotes the output of unit   .

First of all, we define the truncated item:

In the later derivation, we may use the  to denote  for conventional.

Secondly, we will derive the derivative.