知识卡片

普通读书笔记卡 · 8942.a · 生成模型:GPT-6

查询键值承担不同角色

理解等级:初级

查询表示当前希望寻找什么,键表示各候选内容如何被匹配,值表示实际被取回的信息。同一输入可经过三组不同投影分别形成它们。用数据库检索作类比有助于分清角色,但神经注意力通常是可微的软汇集,不是精确查表。

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

      Transformers use a particular kind of attention based on the idea of queries, keys, and values. The idea of queries, keys, and values comes from databases, where a database cell holds a value, which is retrieved when a query matches the cell's key. Tokens are like database cells and attention is like retrieving information from the database of tokens. In query-key-value attention, each token is

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第26章《Transformer》(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/c026.xhtml) - 结论依据:摘录中关于查询键值承担不同角色的定义、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:Transformers use a particular kind of attention based on the idea of queries, keys, and values. The idea of queries, keys, and values comes from databases, where a database cell holds a value, which is retrieved when a query matches the cell's key. Tokens are like database cells and attention is like retrieving information from the database of tokens. In query-key-value attention, each token is