知识卡片

结构图卡 · 8873.g · 生成模型:GPT-6

反向传播的一次求导流程

理解等级:中级

先确定本次工作点,再复用局部导数逐层回传;遇到分支和共享量时累计贡献,最后交给优化器更新。图示不将反传当作前向逆运算。

结构图:

flowchart TD
 A[当前参数与输入] --> B[按图前向并保存所需值]
 B --> C[标量损失及初始导数1]
 C --> D[局部向量雅可比积]
 D --> E[累加分支及共享参数贡献]
 E --> F{还有上游依赖}
 F -->|有| D
 F -->|无| G[输入梯度与参数梯度]
 G --> H[优化器按约束更新自由变量]
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

      The full algorithm is summarized in algorithm 14.1. Algorithm 14.1: Backpropagation (for chain computation graphs). A simple version of the backpropagation algorithm. This will work for the computation graphs we have seen so far, which consist of a series of layers, f1 ◦ … ◦ fL, with no merging or branching (see section 14.7 for how to handle morecomplicated graphs with merge and branch operations).

    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:摘录中关于反向传播的一次求导流程的定义、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:The full algorithm is summarized in algorithm 14.1. Algorithm 14.1: Backpropagation (for chain computation graphs). A simple version of the backpropagation algorithm. This will work for the computation graphs we have seen so far, which consist of a series of layers, f1 ◦ … ◦ fL, with no merging or branching (see section 14.7 for how to handle morecomplicated graphs with merge and branch operations).