知识卡片
双线性匹配不是距离度量
查询键内积可以改写为tᵢᵀS tⱼ,其中S=WqᵀWk。S不必对称或半正定,所以该分数不必是合法距离或对称相似度。只有适当归一化后,点积才可按余弦相似度解释,不能对任意投影直接这样命名。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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
What this shows is that Wq and Wk define some matrix that modulates how we measure similarity (dot product) between ti and tj. A single self-attention layer therefore measures similarity in just one way.
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:摘录中关于双线性匹配不是距离度量的定义、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:What this shows is that Wq and Wk define some matrix that modulates how we measure similarity (dot product) between ti and tj. A single self-attention layer therefore measures similarity in just one way.