知识卡片
逐元素裁剪不同于范数裁剪
原书算法将每个梯度分量限制在[−m,m],这是逐元素裁剪;整个向量范数仍可能随维度增大。按整体范数缩放则能保留非零向量方向,两者不可混称。裁剪有助限制过大更新,却不能自动修复NaN或不存在的导数。
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
What about figure 10.3(e), where the gradient explodes near the optimum? Is there anything we can do to improve optimization of this function? To combat exploding gradients, a useful trick is gradient clipping, which just means clamping the magnitude of the gradient to some maximum value. Algorithm 10.5 describes this approach.
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:摘录中关于逐元素裁剪不同于范数裁剪的定义、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:What about figure 10.3(e), where the gradient explodes near the optimum? Is there anything we can do to improve optimization of this function? To combat exploding gradients, a useful trick is gradient clipping, which just means clamping the magnitude of the gradient to some maximum value. Algorithm 10.5 describes this approach.