知识卡片
长雅可比链导致梯度失衡
跨T步的导数包含多个状态雅可比矩阵连乘。ReLU简单RNN的局部雅可比为RₜW,其中Rₜ由当前激活门控决定;连乘可能在某些方向放大、另一些方向衰减。不能仅凭W各元素看起来大或小判断,矩阵谱、奇异值、非正规性和门控路径都重要。
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
- · 循环神经网络
原书出处 · 1 处
- · 查看原书依据
Notice that the we can get very large computation graphs when we run an RNN for many timesteps. The gradient computation involves a series of matrix multiplies that is O(T) in length. For example, to compute our gradient in the example above involves the following product:
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第25章《循环神经网络》(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/c025.xhtml)
- 结论依据:摘录中关于长雅可比链导致梯度失衡的定义、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:Notice that the we can get very large computation graphs when we run an RNN for many timesteps. The gradient computation involves a series of matrix multiplies that is O(T) in length. For example, to compute our gradient in the example above involves the following product: