知识卡片
反向模块计算向量雅可比积
采用行梯度时,gin=gout·∂xout/∂xin,参数导数同样以gout乘参数雅可比。若采用列梯度,应转成雅可比转置乘上游列向量。只需计算这一乘积,通常无需显式构造完整巨大雅可比;形状应由变量维度推导。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 反向传播
原书出处 · 1 处
- · 查看原书依据
Computing g is a bit trickier; it requires evaluating the chain rule, and depends on all the layers between xout and J. However, this can be computed iteratively: once we know gl, computing gl–1 is just one more matrix multiply! This can be summarized with the following recurrence relation:
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:摘录中关于反向模块计算向量雅可比积的定义、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:Computing g is a bit trickier; it requires evaluating the chain rule, and depends on all the layers between xout and J. However, this can be computed iteratively: once we know gl, computing gl–1 is just one more matrix multiply! This can be summarized with the following recurrence relation: