← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 9173.d · 生成模型:GPT-6

最大池化反传路由到获胜位置

理解等级:中级

窗口有唯一最大值时,上游梯度传到该位置,其他位置为零;并列最大需实现约定。平均池化则按窗口规则分摊梯度。分段可导操作可用于反传,并不要求所有层处处光滑。

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

      Although the max() function is not differentiable, MaxPooling can still be used in backpropagation by selectively passing the gradient to only the neuron that has shown to have the maximum activation and setting the gradient of all other neurons to zero. When an averaging pooling function is used, the gradient is divided among all neurons in the pool in equal parts.

    参考来源

    - 位置:《Introduction to Autonomous Robots Mechanisms, Sensors, Actuators, and Algorithms (Nikolaus Correll, Bradley Hayes etc.)》第10章《人工神经网络》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Introduction to Autonomous Robots Mechanisms, Sensors, Actuators, and Algorithms (Nikolaus Correll, Bradley Hayes etc.)/_epub-src/OEBPS/xhtml/chapter_10.xhtml) - 结论依据:摘录中关于最大池化反传路由到获胜位置的定义、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:Although the max() function is not differentiable, MaxPooling can still be used in backpropagation by selectively passing the gradient to only the neuron that has shown to have the maximum activation and setting the gradient of all other neurons to zero. When an averaging pooling function is used, the gradient is divided among all neurons in the pool in equal parts.