知识卡片
全局注意力承担平方成本
N个查询各与N个键匹配,标准全局注意力需处理N×N分数或等价计算,通常带来平方级序列成本。共享参数减少的是独立权重数量,不自动消除该成本。计算效率需要结合词元数、维度、实现和所需交互范围讨论。
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 are not convolutional. They do not factor the processing into independent chunks but instead perform a global operation, in which all input tokens can interact. The linear combination of tokens that results is not a mixing operation found in CNNs and addresses the limitation of CNNs being myopic, with each filter only seeing information in its receptive field.
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:摘录中关于全局注意力承担平方成本的定义、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 are not convolutional. They do not factor the processing into independent chunks but instead perform a global operation, in which all input tokens can interact. The linear combination of tokens that results is not a mixing operation found in CNNs and addresses the limitation of CNNs being myopic, with each filter only seeing information in its receptive field.