← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8856.d · 生成模型:GPT-6

扰动采样不是保证每步改善

理解等级:中级

随机扰动可以跨越局部零梯度区,但有限样本仍有噪声,未必每步降低原损失。原书图10.4实际对应分段常数的图10.3(d),正文误指©。算法每轮也应围绕当前θₖ采样,不能把伪代码未标上标的θ误当作始终固定初始参数。

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

      get v is to compute it by sampling perturbations of θ, and seeing which perturbation leads to lower loss. In this strategy, we evaluate J(θ + ϵ) for a set of perturbations ϵ, then move toward the ϵ's that decreased the loss. Approaches of this kind are sometimes called evolution strategies [47, 421], and a basic version of this algorithm is given in algorithm 10.4.

    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:摘录中关于扰动采样不是保证每步改善的定义、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:get v is to compute it by sampling perturbations of θ, and seeing which perturbation leads to lower loss. In this strategy, we evaluate J(θ + ϵ) for a set of perturbations ϵ, then move toward the ϵ's that decreased the loss. Approaches of this kind are sometimes called evolution strategies [47, 421], and a basic version of this algorithm is given in algorithm 10.4.