← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8856 · 生成模型:GPT-6

梯度消失与爆炸阻碍不同

理解等级:中级

梯度接近零时,固定学习率更新很小,模型可能停滞;梯度非常大时,步长可能失控并越过最优区域。二者都是下降信号尺度不良,却需不同处理。损失连续或训练曲线暂时平稳,并不自动排除这些问题。

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

      we have a gradient that has nearly vanished, that is, it is near zero everywhere and gradient descent, with a fixed learning rate, will therefore be slow. Figure 10.3(e) shows the opposite scenario: the gradient at the minimizer goes to infinity; we call this an exploding gradient, and this leads to failures of convergence.

    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:摘录中关于梯度消失与爆炸阻碍不同的定义、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:we have a gradient that has nearly vanished, that is, it is near zero everywhere and gradient descent, with a fixed learning rate, will therefore be slow. Figure 10.3(e) shows the opposite scenario: the gradient at the minimizer goes to infinity; we call this an exploding gradient, and this leads to failures of convergence.