知识卡片
学习到的关联不保证解释
注意力可将相似部位的信息汇集起来,书中动物颜色的例子说明这种可能用途。但头的职责由训练形成,不保证某个头始终代表颜色或对象身份;高权重表示这一计算中的汇集强度,不能直接等同因果解释或可靠语义归因。
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
This is just one way self-attention could be used by the network. How it is actually used will be determined by the training data and task. What really happens might deviate from our intuitive story: tokens on hidden layers do not necessarily represent spatially localized patches of pixels. While the initial tokenization layer creates tokens out of local image patches, after this point attention layers
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:摘录中关于学习到的关联不保证解释的定义、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:This is just one way self-attention could be used by the network. How it is actually used will be determined by the training data and task. What really happens might deviate from our intuitive story: tokens on hidden layers do not necessarily represent spatially localized patches of pixels. While the initial tokenization layer creates tokens out of local image patches, after this point attention layers