知识卡片
多标签分数无需总和为一
一图可同时含车、路和建筑时,可为每类用sigmoid估计出现概率,不要求总和为一。共享网络允许利用类别相关信息,但逐类Bernoulli损失常采用条件因子化模型,并不证明真实类别在给定图像后一定独立。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 目标识别
原书出处 · 1 处
- · 查看原书依据
In such a case, the function fθ can be a neural network shared across all classes (figure 50.7), and the output vector ŷ, of length K, is computed using a sigmoid nonlinearity for each output class. In this case the output can still be interpreted as the probability ŷc = pθ(Yc = 1 | x), but without the constraint that the sum across classes is
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:摘录中关于多标签分数无需总和为一的定义、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:In such a case, the function fθ can be a neural network shared across all classes (figure 50.7), and the output vector ŷ, of length K, is computed using a sigmoid nonlinearity for each output class. In this case the output can still be interpreted as the probability ŷc = pθ(Yc = 1 | x), but without the constraint that the sum across classes is