知识卡片

普通读书笔记卡 · 6743 · 生成模型:GPT-6

注意力权重是输入相关混合

理解等级:中级

输入分别投影为Q、K、V,以softmax(QKᵀ/√d_k)得到可见键的权重,再对V聚合。权重随输入改变,整体不只是一个固定Linear层;方阵也不保证正交或保范数,不能把注意力简单称为旋转。

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

      Self-attention is the most straightforward and common way to implement attention. It takes the input sequence of embedding vectors and puts them through linear projections. A linear projection is merely a dot product or matrix multiplication. This dot product creates key, value, and query vectors. The query vector is used along with the key vector to create a context vector for the words’ embedding vectors

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(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-9.html) - 结论依据:摘录中关于注意力权重是输入相关混合的定义、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:Self-attention is the most straightforward and common way to implement attention. It takes the input sequence of embedding vectors and puts them through linear projections. A linear projection is merely a dot product or matrix multiplication. This dot product creates key, value, and query vectors. The query vector is used along with the key vector to create a context vector for the words’ embedding vectors