知识卡片

普通读书笔记卡 · 8871.b · 生成模型:GPT-6

仿射层把梯度乘回权重转置

理解等级:中级

对y=Wx+b,采用与变量同形的列梯度δy,输入梯度为δx=Wᵀδy。偏置不出现在这一输入雅可比里,但仍通过前向值影响后续梯度。Wᵀ是链式求导,不是W的逆矩阵,也不保证由输出梯度恢复前向输入。

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

      backward for a linear layer is the same operation as forward, just with the weights transposed! We have omitted the bias terms here, but recall from equation (14.18) that the backward pass to the activations ignores biases anyway.

    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:摘录中关于仿射层把梯度乘回权重转置的定义、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:backward for a linear layer is the same operation as forward, just with the weights transposed! We have omitted the bias terms here, but recall from equation (14.18) that the backward pass to the activations ignores biases anyway.