← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8930.b · 生成模型:GPT-6

跨通道池化需有对应语义

理解等级:中级

若不同通道分别检测同一结构的不同方向,跨通道max可以汇总“任一方向出现”的证据。它失去被选中方向的信息,且只覆盖已有通道检测的变换;普通学习通道未必天然构成完整旋转族,因此不能只加通道池化就断言模型旋转不变。

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

      Pooling can also be performed across channels, and this can be a way to achieve additional kinds of invariance. For example, suppose we have a convolutional layer that applies a filter bank of oriented edge detector filters, where each filter looks for edges at a different orientation. Now if we max pool across the channels output by this filter bank, the resulting feature map will

    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:摘录中关于跨通道池化需有对应语义的定义、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:Pooling can also be performed across channels, and this can be a way to achieve additional kinds of invariance. For example, suppose we have a convolutional layer that applies a filter bank of oriented edge detector filters, where each filter looks for edges at a different orientation. Now if we max pool across the channels output by this filter bank, the resulting feature map will