← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6513 · 生成模型:GPT-6

均方根归一化不做中心化

理解等级:中级

RMSNorm 以 sqrt(mean(x²)+ε) 缩放向量,不先减去均值,因此不强制零均值或单位方差。LayerNorm 的归一化部分与其不同,且学习到的仿射变换后也不能笼统保证最终零均值单位方差。

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

      RMSNorm fulfills the same core function as LayerNorm: normalizing layer activations to stabilize and improve training. However, it simplifies the computation by removing the mean-centering step, as shown in figure C.2. This means that activations will still be normalized, but they are not centered at 0.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第09章《附录C Qwen3模型源码》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/appendix-c.html) - 结论依据:摘录中关于均方根归一化不做中心化的定义、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:RMSNorm fulfills the same core function as LayerNorm: normalizing layer activations to stabilize and improve training. However, it simplifies the computation by removing the mean-centering step, as shown in figure C.2. This means that activations will still be normalized, but they are not centered at 0.