知识卡片

普通读书笔记卡 · 6762.a · 生成模型:GPT-6

未归一分数不是概率的唯一逆

理解等级:中级

logits是词表候选的未归一分数,经softmax才形成分布;给全部分数加同一常数,概率不变,所以softmax没有唯一逆。二元logit的定义又不同,不能把模型输出logits直接叫概率或softmax的唯一反函数。

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

      That’s an interesting type for the output! If you look at the Hugging Face documentation,38 you’ll see that it has a lot of interesting information inside—from the hidden states of the model to attention weights for self-attention and cross-attention. What we’re going to look at, however, is the part of the dictionary called logits. The logit function is the inverse of the softmax function—it maps

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第10章《真实世界中的大语言模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-10.html) - 结论依据:摘录中关于未归一分数不是概率的唯一逆的定义、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:That’s an interesting type for the output! If you look at the Hugging Face documentation,38 you’ll see that it has a lot of interesting information inside—from the hidden states of the model to attention weights for self-attention and cross-attention. What we’re going to look at, however, is the part of the dictionary called logits. The logit function is the inverse of the softmax function—it maps