← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8873.a · 生成模型:GPT-6

复制分支的梯度必须相加

理解等级:中级

同一变量复制到多个下游路径后,最终损失会沿每条路径依赖它。反传时应将全部路径贡献相加,再继续传给上游。覆盖其中一条或自行取平均都会改变导数;求平均仅在目标本身包含相应缩放时出现。

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, branching just sums both the gradients passed backward to it. Both merge and branch have no parameters, so there is no parameter gradient to define. Thus, we have fully specified the forward and backward behavior of these layers. The next diagrams summarize the behavior (figure 14.19). Figure 14.19: Merge and branch layers forward and backward. With merge and branch, we can construct any DAG

    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:摘录中关于复制分支的梯度必须相加的定义、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, branching just sums both the gradients passed backward to it. Both merge and branch have no parameters, so there is no parameter gradient to define. Thus, we have fully specified the forward and backward behavior of these layers. The next diagrams summarize the behavior (figure 14.19). Figure 14.19: Merge and branch layers forward and backward. With merge and branch, we can construct any DAG