知识卡片

普通读书笔记卡 · 6738 · 生成模型:GPT-6

重置门控制候选中的旧状态贡献

理解等级:中级

GRU用reset门r调节旧状态对候选n的影响,用update门z在候选与旧状态之间混合。门通常由sigmoid产生连续值,不是硬if,原书只列门值不等于已给出完整新状态更新。

这张卡片看不懂?

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

      In addition to the original RNN output gate, GRUs add two new logic gates or activation functions within your recurrent unit: Reset gates—What parts of the hidden layer should be blocked because they are no longer relevant to the current output? Update gates—What parts of the hidden layer should matter to the current output (now, at time t)? You already had an activation function on

    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:摘录中关于重置门控制候选中的旧状态贡献的定义、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:In addition to the original RNN output gate, GRUs add two new logic gates or activation functions within your recurrent unit: Reset gates—What parts of the hidden layer should be blocked because they are no longer relevant to the current output? Update gates—What parts of the hidden layer should matter to the current output (now, at time t)? You already had an activation function on