知识卡片

普通读书笔记卡 · 6509.c · 生成模型:GPT-6

长序列乘积容易数值下溢

理解等级:中级

多个小于一的条件概率相乘,序列稍长就可能接近浮点表示下限。用对数把乘积变成求和,可保留更稳定的数值尺度;数学上对数单调,固定条件下更大的序列概率对应更大的对数概率。

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

      These results make sense since the "Berlin" answer receives a higher sequence likelihood than the "Bridge" answer, as expected. However, both values are extremely small because multiplying many probabilities less than one quickly produces numbers that approach zero. This makes raw likelihoods difficult to work with in practice, especially for longer sequences where underflow becomes a problem.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第05章《自我修正与推理时扩展》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-5.html) - 结论依据:摘录中关于长序列乘积容易数值下溢的定义、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:These results make sense since the "Berlin" answer receives a higher sequence likelihood than the "Bridge" answer, as expected. However, both values are extremely small because multiplying many probabilities less than one quickly produces numbers that approach zero. This makes raw likelihoods difficult to work with in practice, especially for longer sequences where underflow becomes a problem.