知识卡片

普通读书笔记卡 · 8926.b · 生成模型:GPT-6

相关与卷积需匹配权重

理解等级:中级

深度网络常用互相关实现“卷积层”,与数学卷积相比差一个空间翻核。在自由学习的核下二者可表达同样的函数族,但复用已有权重或手工导数核时必须同步翻转,否则响应方向或符号可能改变。等价的假设空间不等于同一数组直接计算相同输出。

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

      “Convolutional” layers in deep nets are typically actually defined as cross-correlations (⋆) and we stick to that convention in this book. We need not worry about the misnomer because whether you implement the layers with convolution or cross-correlation usually makes no difference for learning. This is because both span an identical hypothesis space (any cross-correlation can be converted to an equivalent convolution by flipping the

    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:摘录中关于相关与卷积需匹配权重的定义、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:“Convolutional” layers in deep nets are typically actually defined as cross-correlations (⋆) and we stick to that convention in this book. We need not worry about the misnomer because whether you implement the layers with convolution or cross-correlation usually makes no difference for learning. This is because both span an identical hypothesis space (any cross-correlation can be converted to an equivalent convolution by flipping the