知识卡片

普通读书笔记卡 · 8933.b · 生成模型:GPT-6

主成分颜色不是固定语义

理解等级:中级

将某层特征通道用PCA压到3维,再映射成RGB,可观察空间组织怎样随深度改变。颜色表示该层方差方向,主成分有符号和基变换歧义,若各层分别拟合,颜色不能跨层直接对照同一语义。可视化也只保留部分方差信息。

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

      Figure 24.25 shows one more way to visualize feature maps. Rather than plotting the channels as a column of grayscale images, we run PCA to reduce the channel dimensionality to 3. Then we can directly render each feature map as a color image, with red showing the first priniciple component of each layer's feature map, green the second, and blue the third. We show this

    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:摘录中关于主成分颜色不是固定语义的定义、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:Figure 24.25 shows one more way to visualize feature maps. Rather than plotting the channels as a column of grayscale images, we run PCA to reduce the channel dimensionality to 3. Then we can directly render each feature map as a color image, with red showing the first priniciple component of each layer's feature map, green the second, and blue the third. We show this