知识卡片

普通读书笔记卡 · 8946.a · 生成模型:GPT-6

置换对称可由矩阵推导

理解等级:高级

对置换矩阵P,有Q′=PQ、K′=PK、V′=PV,分数及权重按A′=PAPᵀ重排,故A′V′=PAV。推导要求掩码和位置处理满足相应变换条件;若只重排内容而固定位置标记或固定因果掩码,一般不再保持任意置换对称。

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

      where permute is a permutation of the order of tokens in Tin (i.e., permutes the rows of the matrix). This means that if you scramble (i.e. permute) the patches in the input image then apply attention, the output will be unchanged up to a permutation of the original output. Since the full transformer architecture is just composition of these two types of layers (plus, potentially,

    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:摘录中关于置换对称可由矩阵推导的定义、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:where permute is a permutation of the order of tokens in Tin (i.e., permutes the rows of the matrix). This means that if you scramble (i.e. permute) the patches in the input image then apply attention, the output will be unchanged up to a permutation of the original output. Since the full transformer architecture is just composition of these two types of layers (plus, potentially,