← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8868.c · 生成模型:GPT-6

ReLU在低维边界聚集概率质量

理解等级:高级

二维ReLU把含负坐标的点映到非负象限的坐标轴或原点;高维中则可能落到非负正交锥的多个面。边界上会积累概率质量,因此输出未必有普通全维连续密度。高维的“边界”并不只是一维坐标轴。

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

      relu layer maps many points to the axes of the positive quadrant. In general, with relu-nets, a lot of data density will build up along these axes, because any point not in the positive quadrant snaps to an axis. This effect becomes exaggerated in real networks with high-dimensional embeddings. In particular, for a width-N layer, the region that is strictly positive occupies only proportion of

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第13章《作为分布变换器的神经网络》(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/c013.xhtml) - 结论依据:摘录中关于ReLU在低维边界聚集概率质量的定义、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:relu layer maps many points to the axes of the positive quadrant. In general, with relu-nets, a lot of data density will build up along these axes, because any point not in the positive quadrant snaps to an axis. This effect becomes exaggerated in real networks with high-dimensional embeddings. In particular, for a width-N layer, the region that is strictly positive occupies only proportion of