知识卡片

普通读书笔记卡 · 8944.c · 生成模型:GPT-6

多头并行学习不同汇集

理解等级:中级

多头注意力并行使用多组查询、键和值投影,让不同子空间计算各自的匹配与汇集。各头可能捕捉不同关系,但没有保证自动获得可命名、互不重复的职责。头数增加与堆叠更多层不同,前者并行读同层输入,后者顺序改写表示。

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

      What if we want to measure similarity in more than one way? For example, maybe we want our net to perform some set of computations based on color similarity, another based on texture similarity, and yet another based on shape similarity? The way transformers can do this is with multihead self-attention (MSA). This method simply consists of running k attention layers in parallel. All these

    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:摘录中关于多头并行学习不同汇集的定义、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 if we want to measure similarity in more than one way? For example, maybe we want our net to perform some set of computations based on color similarity, another based on texture similarity, and yet another based on shape similarity? The way transformers can do this is with multihead self-attention (MSA). This method simply consists of running k attention layers in parallel. All these