← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8872.a · 生成模型:GPT-6

平方损失的系数决定梯度尺度

理解等级:中级

若J=‖ŷ−y‖²,则对预测的同形梯度为2(ŷ−y)。若定义中有1/2、按维度平均或按批平均,梯度也随之缩放。原书行列表示需转置,但差值方向不变;更新前应核对损失定义,避免把平方和与均方误差混用。

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, the backward signal sent by the L2 loss layer is a row vector of per-dimension errors between the prediction and the target. This completes our derivation of forward and backward for a L2 loss layer, yielding figure 14.14. Figure 14.14: L2 loss layer forward and backward 14.6.4 Putting It All Together: Backpropagation through an MLP Let's see what happens when we put 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:摘录中关于平方损失的系数决定梯度尺度的定义、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, the backward signal sent by the L2 loss layer is a row vector of per-dimension errors between the prediction and the target. This completes our derivation of forward and backward for a L2 loss layer, yielding figure 14.14. Figure 14.14: L2 loss layer forward and backward 14.6.4 Putting It All Together: Backpropagation through an MLP Let's see what happens when we put all these