← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8975.b · 生成模型:GPT-6

增强选择改变嵌入偏好

理解等级:中级

书中玩具实验用裁剪得到偏颜色的表示,进一步加色相、亮度与饱和度变化后得到偏形状的表示。它说明视图构造可以改变保留信息;裁剪本身并不一般保证形状不变,颜色扰动也不保证只删除颜色而不影响其他任务线索。

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

      Suppose we wish to obtain an embedding that is sensitive to color and invariant to shape. Then we should choose a view transformation Tc that preserves color while changing shape. A simple choice that turns out to work is for Tc(x) to simply output a crop from the image x (this transformation does not really change the object's shape but still ends up resulting in

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第30章《表示学习》(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/c030b.xhtml) - 结论依据:摘录中关于增强选择改变嵌入偏好的定义、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:Suppose we wish to obtain an embedding that is sensitive to color and invariant to shape. Then we should choose a view transformation Tc that preserves color while changing shape. A simple choice that turns out to work is for Tc(x) to simply output a crop from the image x (this transformation does not really change the object's shape but still ends up resulting in