知识卡片

普通读书笔记卡 · 8943 · 生成模型:GPT-6

注意力输出必须加权求和

理解等级:中级

一个查询的输出是Σⱼaⱼvⱼ,将所有候选值按匹配权重相加。把aⱼvⱼ依次堆叠会保留候选轴,形状和含义都不同;书中式26.16的堆叠式应结合正文及矩阵形式理解为加权求和。

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

      v1 is the value vector for t1 =Tin[0, :], and so forth. Figure 26.8 visualizes these steps. We use the following color scheme here and later in this chapter: Figure 26.8: Mechanics of an attention layer. Queries from the question match keys from the tokens representing the impala; value vectors of the impala tokens then contribute the most to the sum that yields tout's code

    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:摘录中关于注意力输出必须加权求和的定义、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:v1 is the value vector for t1 =Tin[0, :], and so forth. Figure 26.8 visualizes these steps. We use the following color scheme here and later in this chapter: Figure 26.8: Mechanics of an attention layer. Queries from the question match keys from the tokens representing the impala; value vectors of the impala tokens then contribute the most to the sum that yields tout's code