知识卡片
分块限制像素平移等变
词元置换等变不能直接推出图像对任意像素平移等变。只有在合适边界及分块对齐条件下,移动图像才可能对应词元置换;跨越小块边界的平移会改变分块内容。绝对位置编码又引入固定坐标,因此需单独检验所需的空间对称性。
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
It is often useful to understand layers in terms of their invariances and equivariances. Convolutational layers are translation equivariant but not necessarily permutation equivariant whereas attention layers are both translation equivariant and permutation equivariant (since translation is a special kind of permutation, any permutation equivariant layer is also translation equivariant). Other layers can be catalogued similarly: global average pooling layers are permutation invariant, relu 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:摘录中关于分块限制像素平移等变的定义、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:It is often useful to understand layers in terms of their invariances and equivariances. Convolutational layers are translation equivariant but not necessarily permutation equivariant whereas attention layers are both translation equivariant and permutation equivariant (since translation is a special kind of permutation, any permutation equivariant layer is also translation equivariant). Other layers can be catalogued similarly: global average pooling layers are permutation invariant, relu layers