知识卡片

普通读书笔记卡 · 8943.a · 生成模型:GPT-6

交叉注意力分开信息来源

理解等级:中级

交叉注意力允许查询来自一组表示,而键和值来自另一组表示,例如用任务查询读取图像内容。查询数决定输出行数,候选键数决定每行匹配范围。它不要求两组序列具有相同长度,但投影后的查询键维度必须相容。

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

      As we have now seen, attention is a general-purpose way of dynamically pooling information in one set of tokens based on queries from a different set of tokens. The next question we will consider is which tokens should be doing the querying and which should we be matching against? In the example from the last section, the answer was intuitive because we had a textual

    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:摘录中关于交叉注意力分开信息来源的定义、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:As we have now seen, attention is a general-purpose way of dynamically pooling information in one set of tokens based on queries from a different set of tokens. The next question we will consider is which tokens should be doing the querying and which should we be matching against? In the example from the last section, the answer was intuitive because we had a textual