知识卡片
结构稀疏不等同低秩
卷积矩阵具有局部、共享的结构限制,但不必低秩,例如恒等卷积可满秩。注意力的QKᵀ秩至多为查询键维度,经过逐元素指数及逐行归一化后,A仍可能满秩。比较架构时应分别讨论参数共享、稀疏性、秩和输入依赖。
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
Many layers in deep nets are special kinds of affine transformations. Three we have seen so far are fully-connected layers (fc), convolutional layers (conv), and self-attention layers (attn). All these layers are alike in that their forward pass can be written as Xout = WXin + b for some matrix W and some vector b. In conv and attn layers, W and b are determined
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:摘录中关于结构稀疏不等同低秩的定义、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:Many layers in deep nets are special kinds of affine transformations. Three we have seen so far are fully-connected layers (fc), convolutional layers (conv), and self-attention layers (attn). All these layers are alike in that their forward pass can be written as Xout = WXin + b for some matrix W and some vector b. In conv and attn layers, W and b are determined