← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8871.a · 生成模型:GPT-6

平均损失的梯度也取平均

理解等级:中级

可微且求和有限时,平均损失的梯度等于各项梯度平均,因此可在批量计算中汇总样本贡献。若层使用批统计或损失含跨样本项,应对真实整批计算图求导,不能假装每个样本都能完全独立前向再简单拼接。

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 gradient of a sum of terms is the sum of the gradients of each term. where Ji(θ) is the loss for a single datapoint x(i). Therefore, to compute a gradient update for an algorithm like stochastic gradient descent (section 10.7), we apply backpropagation in batch mode, that is, we run it over each datapoint in our batch (which can be done in parallel) and

    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:摘录中关于平均损失的梯度也取平均的定义、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 gradient of a sum of terms is the sum of the gradients of each term. where Ji(θ) is the loss for a single datapoint x(i). Therefore, to compute a gradient update for an algorithm like stochastic gradient descent (section 10.7), we apply backpropagation in batch mode, that is, we run it over each datapoint in our batch (which can be done in parallel) and