deep-learning - how does LSTM and GRU gates decide which word to keep in the memory

Question

the update gate in a GRU decides which word to keep in the cell or to be clear what is the cell state. how does the update gate in gru decide when to be close to 1 and when to be close to 0? Basically, how does it decide to keep a word and not to keep a word? Thanks

Answer 1

目前尚不清楚您在该问题中做了哪些背景研究。也许谷歌搜索 gru 教程可能会出现一些有用的博客。这是一个。

请记住，尽管单元状态是使用权重进行更新的，这些权重的优化方式与任何其他神经网络权重相同，具有反向传播。这是一篇很棒的博客，可以帮助解释如何设置 RNN（GRU 是一个特定示例）以及如何使用反向传播来更新权重。然后 GRU 的权重决定了如何更新单元状态。