知识卡片
梯度优化的选择与诊断
先识别可用局部信号,再选择更新与稳定化机制。调度、动量、批量和裁剪解决不同问题,不应只凭一个训练曲线标签机械叠加。
结构图:
flowchart TD
A[目标与当前参数] --> B{梯度是否有效}
B -->|有| C[全量或小批量估计]
B -->|缺乏有效信号| D[替代损失或扰动搜索]
C --> E[学习率与动量]
E --> F[必要时裁剪过大梯度]
D --> G[更新参数并检查原目标]
F --> G
G --> A
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
原书出处 · 1 处
- · 查看原书依据
or not, the handful of algorithms described above are enough to train most state-of-the-art deep learning models. Every year there are new elaborations on these ideas, and second-order methods are ever on the horizon, yet the basic concepts remain quite simple: compute a local estimate of the shape of the loss landscape, then, based on this shape, take a small step toward a lower loss.
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:摘录中关于梯度优化的选择与诊断的定义、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:or not, the handful of algorithms described above are enough to train most state-of-the-art deep learning models. Every year there are new elaborations on these ideas, and second-order methods are ever on the horizon, yet the basic concepts remain quite simple: compute a local estimate of the shape of the loss landscape, then, based on this shape, take a small step toward a lower loss.