← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8871.c · 生成模型:GPT-6

权重梯度是误差信号与输入外积

理解等级:高级

若W形状为M×N、输入x为N维列向量、上游列梯度δy为M维,则同形权重梯度为δy xᵀ,偏置梯度为δy。原书按矩阵导数布局写x gout,得到N×M后需转置更新;不能把它直接当W同形梯度使用。

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

      So we see that in the end this gradient has the simple form of an outer product between two vectors, xin and gout (figure 14.9). Figure 14.9: Matrix multiply for parameter gradient of a linear layer. We can summarize all these operations in the forward and backward diagram for linear layer in figure 14.10. Figure 14.10: Linear layer forward and bacward. Notice that all these

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第14章《反向传播》(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/c014.xhtml) - 结论依据:摘录中关于权重梯度是误差信号与输入外积的定义、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:So we see that in the end this gradient has the simple form of an outer product between two vectors, xin and gout (figure 14.9). Figure 14.9: Matrix multiply for parameter gradient of a linear layer. We can summarize all these operations in the forward and backward diagram for linear layer in figure 14.10. Figure 14.10: Linear layer forward and bacward. Notice that all these