知识卡片

普通读书笔记卡 · 9085.d · 生成模型:GPT-6

二元交叉熵包含正负两项

理解等级:中级

多标签分类用−Σ_c[y_c log p_c+(1−y_c)log(1−p_c)],既奖励真类出现,也约束未出现类。书中式50.10排版的负号作用范围不清,必须覆盖整个括号;只保留正类项会丢掉大量错误阳性约束。

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 the outputs ŷc are computed using a sigmoid layer. Taking the negative log of the likelihood gives the multiclass binary cross-entropy loss: (50.10) 50.3.3 Evaluation Once the function fθ has been trained, we have to evaluate its performance over a holdout test set. There are several popular measures of performance. Classification performance uses the output of the classifier, f, as a continuous value, 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:摘录中关于二元交叉熵包含正负两项的定义、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 the outputs ŷc are computed using a sigmoid layer. Taking the negative log of the likelihood gives the multiclass binary cross-entropy loss: (50.10) 50.3.3 Evaluation Once the function fθ has been trained, we have to evaluate its performance over a holdout test set. There are several popular measures of performance. Classification performance uses the output of the classifier, f, as a continuous value, the