知识卡片
两种混合沿不同轴进行
词元矩阵T的行为位置、列为特征。左乘位置混合矩阵使一个输出词元汇集多个输入位置;右乘共享特征投影则在各词元内部重组特征。区分两个轴,才能看清注意力负责交换位置间信息,而逐词元网络负责处理已汇集的信息。
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
Transformers consist of two main operations over tokens: (1) mixing tokens via a weighted sum, and (2) modifying each individual token via a nonlinear transformation. These operations are analogous to the two workhorses of regular neural nets: the linear layer and the pointwise nonlinearity.
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:摘录中关于两种混合沿不同轴进行的定义、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:Transformers consist of two main operations over tokens: (1) mixing tokens via a weighted sum, and (2) modifying each individual token via a nonlinear transformation. These operations are analogous to the two workhorses of regular neural nets: the linear layer and the pointwise nonlinearity.