美文网首页
Lecture 13.5 | (2.5/5) Recurrent

Lecture 13.5 | (2.5/5) Recurrent

作者: Ysgc | 来源:发表于2019-11-02 10:21 被阅读0次
require the Network to remember info arbitrarily long

\sigma -> continue to remember or forget it -> a switch

more complex, but the crux is the same: gate looks at input and decide to remember or forget

decide the input to influence the memory

hidden state is not exactly the memory itself, it's the function of the memory.

straightforward

high level why separate the memory and hidden layer??? and input gate + forget gate = 1

GRU is even better than LSTM in some cases, and more follow the rationale of the original idea

相关文章

网友评论

      本文标题:Lecture 13.5 | (2.5/5) Recurrent

      本文链接:https://www.haomeiwen.com/subject/nwscbctx.html