← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8929 · 生成模型:GPT-6

稀疏参数不等于低矩阵秩

理解等级:中级

空洞核在更大支撑中只保留少量非零位置,因此是稀疏参数化;其二维核矩阵仍可能满秩。低秩空间核则要求能用少量行列外积表示。两种方法都可能节约参数,却施加不同约束,不能仅凭权重少就称核矩阵低秩。

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

      Dilation is one way to create a big filter that is parameterized by just a small number of weights, that is, a low-rank filter. This trick can be useful in many contexts where we know that good filters have low-rank structure. Dilation uses this trick to make big kernels, which can capture long-range dependences.

    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:摘录中关于稀疏参数不等于低矩阵秩的定义、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:Dilation is one way to create a big filter that is parameterized by just a small number of weights, that is, a low-rank filter. This trick can be useful in many contexts where we know that good filters have low-rank structure. Dilation uses this trick to make big kernels, which can capture long-range dependences.