知识卡片

普通读书笔记卡 · 8965.c · 生成模型:GPT-6

消息缩放改善数值稳定

理解等级:中级

消息乘一个与其状态无关的正数,不改变最终归一化边缘,因此可把每条消息归一到和为1或最大值为1,以减少下溢。若所有状态支持为零,应检查矛盾证据或模型约束。需要证据概率或配分函数时,还要保留这些缩放常数。

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

      For numerical stability, we typically normalize the computed messages in equation (29.14) so the entries sum to 1, or so their maximum entry is 1, then remember to renormalize the final marginal probabilities to sum to 1. Here, we will normalize the messages for simplicity, absorbing the normalization into constants, ki.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第29章《概率图模型》(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/c029.xhtml) - 结论依据:摘录中关于消息缩放改善数值稳定的定义、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:For numerical stability, we typically normalize the computed messages in equation (29.14) so the entries sum to 1, or so their maximum entry is 1, then remember to renormalize the final marginal probabilities to sum to 1. Here, we will normalize the messages for simplicity, absorbing the normalization into constants, ki.