知识卡片
自注意与交叉注意的两端不同
自注意力的Q、K、V来自同一序列,交叉注意力可让目标表示作Q、源编码作K和V。后者权重矩阵可为目标长×源长而非方阵;核对来源有助区分可视化究竟显示哪种联系。
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
Unlike RNNs, where there is recurrence and shared weights, in self-attention, all of the vectors used in the query, key, and value matrices come from the input sequences’ embedding vectors. The entire mechanism can be implemented with highly optimized matrix multiplication operations. And the Q × K product forms a square matrix that can be understood as the connection between words in the input sequence.
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:摘录中关于自注意与交叉注意的两端不同的定义、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:Unlike RNNs, where there is recurrence and shared weights, in self-attention, all of the vectors used in the query, key, and value matrices come from the input sequences’ embedding vectors. The entire mechanism can be implemented with highly optimized matrix multiplication operations. And the Q × K product forms a square matrix that can be understood as the connection between words in the input sequence.