知识卡片

普通读书笔记卡 · 9220 · 生成模型:GPT-6

反向传播是链式法则的高效组织

理解等级:中级

从标量损失逆着计算图传播局部导数,可复用中间结果并得到许多参数的梯度。它负责求导,不等于梯度下降本身;反向或前向模式的效率取决于输入输出维数,不能无条件说反向总更快。

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

      Applying the chain rule can be done in two ways: moving forward or backward through the computation graph. Actually doing this by hand for a simple graph shows that going backward is significantly more efficient. Manually deriving the individual partial derivatives also illustrates that many of the computations can actually be recycled. This solution is known as backpropagation (Werbos 1990), a technique that has been

    参考来源

    - 位置:《Introduction to Autonomous Robots Mechanisms, Sensors, Actuators, and Algorithms (Nikolaus Correll, Bradley Hayes etc.)》附录D《附录D:反向传播》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Introduction to Autonomous Robots Mechanisms, Sensors, Actuators, and Algorithms (Nikolaus Correll, Bradley Hayes etc.)/_epub-src/OEBPS/xhtml/appendixD.xhtml) - 结论依据:摘录中关于反向传播是链式法则的高效组织的定义、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:Applying the chain rule can be done in two ways: moving forward or backward through the computation graph. Actually doing this by hand for a simple graph shows that going backward is significantly more efficient. Manually deriving the individual partial derivatives also illustrates that many of the computations can actually be recycled. This solution is known as backpropagation (Werbos 1990), a technique that has been