知识卡片

普通读书笔记卡 · 8948.a · 生成模型:GPT-6

位置编码引入空间顺序

理解等级:中级

没有位置信息时,词元内容的任意置换无法表达固定布局的差异。位置编码把坐标信息注入表示,使模型能够区分相同内容出现于不同位置。书中采用拼接示意,其他结构也可采用相加;具体设计决定维度及对称性如何变化。

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

      Another idea associated with transformers is positional encoding. Operations over tokens in a transformer are permutation equivariant, which means that we can shuffle the positions of the tokens and nothing substantial changes (the only change is that the outputs get permuted). A consequence is that tokens do not know encode their position within the representation of the signal.

    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:摘录中关于位置编码引入空间顺序的定义、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:Another idea associated with transformers is positional encoding. Operations over tokens in a transformer are permutation equivariant, which means that we can shuffle the positions of the tokens and nothing substantial changes (the only change is that the outputs get permuted). A consequence is that tokens do not know encode their position within the representation of the signal.