← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8867 · 生成模型:GPT-6

单位范数化与标准化并不相同

理解等级:高级

L2归一化把非零向量除以其范数,得到单位长度,但不要求零均值。层标准化先中心化再除以标准差:忽略稳定项且方差非零时,其长度为√C而非1;随后逐元素仿射还会改变均值和长度。零向量与零方差均需数值处理。

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

      There are numerous other normalization layers that have been defined over the years. Two more that we will highlight are L2 normalization and layer normalization (layernorm) [26]. L2 normalization projects the inputs onto the unit hypersphere, which useful for bounding the activations to unit vectors:

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第12章《神经网络》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c012.xhtml) - 结论依据:摘录中关于单位范数化与标准化并不相同的定义、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:There are numerous other normalization layers that have been defined over the years. Two more that we will highlight are L2 normalization and layer normalization (layernorm) [26]. L2 normalization projects the inputs onto the unit hypersphere, which useful for bounding the activations to unit vectors: