知识卡片
自注意力从同组词元取值
自注意力从同一词元矩阵生成Q、K、V,使每个位置查询这组表示中的其他位置。三者共享输入来源,不代表共享参数或数值相同;独立投影允许“寻找什么”“如何被找到”和“提供什么内容”分开学习。
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
Self-attention is just such an architecture. The idea is that on a self-attention layer, all tokens submit queries, and for each of these queries, we take a weighted sum over all tokens in that layer. If Tin is a set of N input tokens, then we have N queries, N weighted sums, and N output tokens to form Tout. This is visualized below in figure
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:摘录中关于自注意力从同组词元取值的定义、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 just such an architecture. The idea is that on a self-attention layer, all tokens submit queries, and for each of these queries, we take a weighted sum over all tokens in that layer. If Tin is a set of N input tokens, then we have N queries, N weighted sums, and N output tokens to form Tout. This is visualized below in figure