知识卡片
细胞状态分开保留与写入路径
LSTM维护c与h,常用c_t=ft⊙c(t−1)+i_t⊙g_t,再由h_t=o_t⊙tanh(c_t)形成可见隐藏状态。f、i、o为sigmoid门,g为候选;它们控制保留、写入与输出,不保证c存全部长记忆、h仅存短记忆。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
8.4.3 Long short-term memory LSTM neurons add two more internal gates, in an attempt to improve both the long-term and the short-term memory capacity of an RNN. The LSTM then adds two new gates: one for forgetting and another to gate the LSTM output, bringing it to a total of four internal gates, each with a specific purpose: Forgetting gate (f)—What elements of the hidden
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-8.html)
- 结论依据:摘录中关于细胞状态分开保留与写入路径的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:8.4.3 Long short-term memory LSTM neurons add two more internal gates, in an attempt to improve both the long-term and the short-term memory capacity of an RNN. The LSTM then adds two new gates: one for forgetting and another to gate the LSTM output, bringing it to a total of four internal gates, each with a specific purpose: Forgetting gate (f)—What elements of the hidden