Attention in transformers, step-by-step | Deep Learning Chapter 6
3Blue1BrownYouTube ↗
延伸阅读 1
字幕解读
一个注意力头先比较查询和键,决定哪些位置要互相影响;真正加到目标表示里的是按这些权重组合的值向量。图里的形容词例子用于解释这两种角色。
作者英文字幕 ↗ · 引文按字幕原文展示,可点时间戳回看片段。
补充字幕原文 2
To measure how well each key matches each query, you compute a dot product between each possible key-query pair.For each column in this diagram, you multiply each of the value vectors by the corresponding weight in that column.