知识卡片

普通读书笔记卡 · 8857.b · 生成模型:GPT-6

批大小平衡估计噪声与成本

理解等级:中级

小批量减少每步处理的数据,通常带来更大的梯度估计噪声;增大批量能改善估计稳定性,但增加单步计算和内存。速度不能仅由样本数量推断,还依赖硬件并行效率。批大小、学习率与训练步数应共同记录和调整。

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

      our estimate of the gradient will be less accurate but faster to compute. Therefore we have a tradeoff between accuracy and speed, and we can navigate this tradeoff with the hyperparameter B. The variant of gradient descent that uses this idea is called stochastic gradient descent (SGD), because each iteration of descent uses a different randomly (stochastically) sampled batch of training data to estimate the

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第10章《基于梯度的学习算法》(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/c010.xhtml) - 结论依据:摘录中关于批大小平衡估计噪声与成本的定义、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:our estimate of the gradient will be less accurate but faster to compute. Therefore we have a tradeoff between accuracy and speed, and we can navigate this tradeoff with the hyperparameter B. The variant of gradient descent that uses this idea is called stochastic gradient descent (SGD), because each iteration of descent uses a different randomly (stochastically) sampled batch of training data to estimate the