知识卡片

普通读书笔记卡 · 8940.b · 生成模型:GPT-6

词元是向量化的信息单位

理解等级:初级

图像词元把局部区域的信息编码成一个向量,后续层以词元为单位混合内容。同一词元内部可以包含多个颜色和形状特征;深层词元已经聚合其他区域的信息,不能始终把它解释为只描述原始小块的局部特征。

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

      We discussed that the main data structures in deep learning are different kinds of groups of neurons: channels, tensors, batches, and so on. Now we will introduce another fundamental data structure, tokens. A token is another kind of group of neurons, but there are particular ways we will operate over tokens that are different from how we operated over channels, batches, and the other groupings

    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:摘录中关于词元是向量化的信息单位的定义、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:We discussed that the main data structures in deep learning are different kinds of groups of neurons: channels, tensors, batches, and so on. Now we will introduce another fundamental data structure, tokens. A token is another kind of group of neurons, but there are particular ways we will operate over tokens that are different from how we operated over channels, batches, and the other groupings