← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8931 · 生成模型:GPT-6

局部归一化规定比较邻域

理解等级:中级

局部归一化用附近空间位置或通道的统计量调整激活,邻域选法与L1/L2/标准化方式产生不同算子。LRN用相邻通道激活平方和构造分母,需明确通道端点和稳定常数。BatchNorm、LayerNorm的统计轴不同,不能仅用“全局归一化”概括其全部行为。

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

      Another kind of nonlinear filter is the local normalization layer. These layers normalize each activation in a feature map by statistics the adjacent activations within some neighborhood. There are many different choices for the type of normalization (L1 norm, L2 norm, standardization, etc.) and many different choices for the shape of the neighborhood, such as a square patch in the spatial dimensions, a set of

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(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/c024b.xhtml) - 结论依据:摘录中关于局部归一化规定比较邻域的定义、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:Another kind of nonlinear filter is the local normalization layer. These layers normalize each activation in a feature map by statistics the adjacent activations within some neighborhood. There are many different choices for the type of normalization (L1 norm, L2 norm, standardization, etc.) and many different choices for the shape of the neighborhood, such as a square patch in the spatial dimensions, a set of