知识卡片
动态权重使整体非线性
自注意力权重A由输入内容计算,输出可写为A(T)V(T)。固定A时,汇集值向量是线性运算;输入改变时A也改变,因此整个自注意力映射通常不是线性或仿射函数。不能根据最后一步矩阵乘法断言整层线性。
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
Attention layers define a special kind of linear combination of tokens. Rather than parameterizing the linear combination with a matrix of free parameters W, attention layers use a different matrix, which we call the attention matrix A. The important difference between A and W is that A is data-dependent, that is, the values of A are a function the data input to the network. In
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:摘录中关于动态权重使整体非线性的定义、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:Attention layers define a special kind of linear combination of tokens. Rather than parameterizing the linear combination with a matrix of free parameters W, attention layers use a different matrix, which we call the attention matrix A. The important difference between A and W is that A is data-dependent, that is, the values of A are a function the data input to the network. In