知识卡片
时间反传复用普通链式法则
BPTT把运行的循环图按时间展开,再从损失反向传播梯度。覆盖完整有限运行、且初始状态按既定方式处理时,可以得到该计算图上的精确梯度。递归并未创造另一套求导规则,难点来自参数重复使用和跨许多步的依赖链。
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
Backpropagation is only defined for DAGs and RNNs are not DAGs. To apply backpropagation to RNNs, we unroll the net for T timesteps, which creates a DAG representation of the truncated RNN, and then do backpropagation through that DAG. This is known as backpropagation through time (BPTT).
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:摘录中关于时间反传复用普通链式法则的定义、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:Backpropagation is only defined for DAGs and RNNs are not DAGs. To apply backpropagation to RNNs, we unroll the net for T timesteps, which creates a DAG representation of the truncated RNN, and then do backpropagation through that DAG. This is known as backpropagation through time (BPTT).