知识卡片

普通读书笔记卡 · 8933.d · 生成模型:GPT-6

感受野与步距要分开递推

理解等级:高级

在一维局部层中,前层感受野跨度r、输入网格步距j,核宽K、扩张d、步长s时,新跨度r′=r+(K−1)dj,新步距j′=sj。纯抽取K=1只扩大相邻输出间隔,不会把单个输出感受野直接乘s。多分支及边界还需分别合并支持域。

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

      Notice that the receptive field grows the deeper we go into the network. To understand why, consider a CNN without nonlinearities. Then the l + 1-th layer is the composition of l convolutional filters. As we saw in section 15.4.1, composing filters results in a new filter with larger support (kernel size). The same happens in a CNN with pointwise nonlinearities, since pointwise operations do

    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:摘录中关于感受野与步距要分开递推的定义、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:Notice that the receptive field grows the deeper we go into the network. To understand why, consider a CNN without nonlinearities. Then the l + 1-th layer is the composition of l convolutional filters. As we saw in section 15.4.1, composing filters results in a new filter with larger support (kernel size). The same happens in a CNN with pointwise nonlinearities, since pointwise operations do