知识卡片
查看核频谱产生可检验假设
把学习到的核同时画在空间与傅里叶域,可发现它们偏好哪些方向和频带。书中两路核更像方向带通,分类头再比较两路响应,这提示网络可能依赖频率分布而非抽象线概念。可视化提供假设,还需用干预输入验证,不能仅凭核形状定论。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 卷积神经网络
原书出处 · 1 处
- · 查看原书依据
What has the network learned? How is it solving the problem? One important part of developing a system is to have tools to prove, understand, and debug it. To understand the network it is useful to visualize the kernels. Figure 24.19 shows the two learned 9 × 9 kernels. The first one looks like an horizontal derivative of a Gaussian filter (as we saw in
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:摘录中关于查看核频谱产生可检验假设的定义、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:What has the network learned? How is it solving the problem? One important part of developing a system is to have tools to prove, understand, and debug it. To understand the network it is useful to visualize the kernels. Figure 24.19 shows the two learned 9 × 9 kernels. The first one looks like an horizontal derivative of a Gaussian filter (as we saw in