← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8929.b · 生成模型:GPT-6

奇异值分解近似非分离核

理解等级:中级

任意M×N空间核可经SVD写成至多min(M,N)个可分离外积之和,按大奇异值保留少数项可得到低秩近似。成本收益取决于核尺寸和保留秩;近似误差也应按任务检查。小核或需要高秩时,分解不一定比直接卷积更省。

这张卡片看不懂?

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

      But we know from chapter 18 that this kernel can be written as a linear combination of two separable kernels: w = Sobelx + Sobely, as defined in equation (18.22). In general, any M × N filter can be decomposed as a linear sum of min(N, M) separable filters. The separable filters can be obtained by applying the singular value decomposition (SVD) to the kernel

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(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/c024b.xhtml) - 结论依据:摘录中关于奇异值分解近似非分离核的定义、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:But we know from chapter 18 that this kernel can be written as a linear combination of two separable kernels: w = Sobelx + Sobely, as defined in equation (18.22). In general, any M × N filter can be decomposed as a linear sum of min(N, M) separable filters. The separable filters can be obtained by applying the singular value decomposition (SVD) to the kernel