知识卡片

普通读书笔记卡 · 9085.c · 生成模型:GPT-6

分类似然导出交叉熵

理解等级:中级

互斥单标签的最大似然等价于最小化−Σ_c y_c log p_c,目标标签为one-hot时只选择真类概率。它是优化用的可微替代,不等于测试零一错误,也不保证较小训练交叉熵一定带来更好跨分布表现。

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

      where is the ground truth value of Yc for image t, and the first product loops over all T training examples while the second product loops over all the classes K. We want to find the parameters θ that maximize the likelihood of the ground truth labels for whole training image. As we have seen in chapter 24, maximizing the likelihood corresponds to minimizing the

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第50章《目标识别》(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/c050b.xhtml) - 结论依据:摘录中关于分类似然导出交叉熵的定义、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:where is the ground truth value of Yc for image t, and the first product loops over all T training examples while the second product loops over all the classes K. We want to find the parameters θ that maximize the likelihood of the ground truth labels for whole training image. As we have seen in chapter 24, maximizing the likelihood corresponds to minimizing the