← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8866.e · 生成模型:GPT-6

层归一化沿单样本特征统计

理解等级:中级

对B×C矩阵,层归一化沿每行特征求均值方差,批归一化在简化例中沿每列的批样本求统计。层归一化不通过批统计耦合不同样本,但并不使原本相关的数据自动满足独立同分布。更高维张量必须另行注明归一化轴。

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

      Layernorm computes a mean and variance over elements of a datapoint xin, and will do so separately for each such datapoint in a batch. Batchnorm computes the mean and variance per channel over datapoints in a batch. If we have a batch stored in the tensor , then what layernorm does looks just like a “transpose” of what batchnorm does. Batchnorm standardizes each element of

    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:摘录中关于层归一化沿单样本特征统计的定义、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:Layernorm computes a mean and variance over elements of a datapoint xin, and will do so separately for each such datapoint in a batch. Batchnorm computes the mean and variance per channel over datapoints in a batch. If we have a batch stored in the tensor , then what layernorm does looks just like a “transpose” of what batchnorm does. Batchnorm standardizes each element of