← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8974.b · 生成模型:GPT-6

成对间隔损失要看符号

理解等级:中级

若D是距离,最小化时正对超出上限应受罚,负对低于下限也应受罚,例如max(D正−m正,0)+max(m负−D负,0)。书中式30.10将第二项写为减号,会奖励近负对,不能按原符号直接实现;具体经典版本还可采用平方等形式。

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

      Here m is a general margin parameter, and mpos and mneg are separate margins for the positive and negative pairs respectively. Notice that the InfoNCE loss is a log softmax over a vector of scores f1(x)Tf2(c)/τ with ; you can therefore think of this loss as corresponding to a classification problem where the ground truth class is x+ and the other possible classes are (refer

    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:摘录中关于成对间隔损失要看符号的定义、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:Here m is a general margin parameter, and mpos and mneg are separate margins for the positive and negative pairs respectively. Notice that the InfoNCE loss is a log softmax over a vector of scores f1(x)Tf2(c)/τ with ; you can therefore think of this loss as corresponding to a classification problem where the ground truth class is x+ and the other possible classes are (refer