池化层的反向传播

神经网络的池化层一般是没有参数更新的，但是它仍旧要参与反向传播的参数传递。那应该怎么传递呢？

前向传播

平均池化和最大池化是两种较为常见的池化方式。先来回顾一下池化层的前向传播方式。
以输入 3x3, 池化核 2x2 为例，（无填充(padding)）则输出为 2x2

平均池化

$$\left[\begin{array}{ccc}{x}_{11}& x_{12}& x_{13}\\ x_{21}& x_{22}& x_{23}\\ x_{31}& x_{32}& x_{33}\end{array}\right]\to \left[\begin{array}{cc}\frac{x_{11}+x_{12}+x_{21}+x_{22}}{4}& \frac{x_{12}+x_{13}+x_{22}+x_{23}}{4}\\ \frac{x_{21}+x_{22}+x_{31}+x_{32}}{4}& \frac{x_{22}+x_{23}+x_{32}+x_{33}}{4}\end{array}\right]$$

最大池化

[ x 11 x 12 x 13 x 21 x 22 x 23 x 31 x 32 x 33 ] → [ m a x { x 11 , x 12 , x 21 , x 22 } m a x { x 12 , x 13 , x 22 , x 23 } m a x { x 21 , x 22 , x 31 , x 32 } m a x { x 22 , x 23 , x 32 , x 33 } ] \left[\begin{matrix} x_{11} & x_{12} & x_{13} \\ x_{21} & x_{22} & x_{23}\\ x_{31} & x_{32} & x_{33} \end{matrix}\right] →\left[\begin{matrix} max\{x_{11},x_{12},x_{21},x_{22}\}& max\{x_{12},x_{13},x_{22},x_{23}\} \\ max\{x_{21},x_{22},x_{31},x_{32}\}& max\{x_{22},x_{23},x_{32},x_{33}\} \end{matrix}\right] ⎣⎡​x11​x21​x31​​x12​x22​x32​​x13​x23​x33​​⎦⎤​→[max{x11​,x12​,x21​,x22​}max{x21​,x22​,x31​,x32​}​max{x12​,x13​,x22​,x23​}max{x22​,x23​,x32​,x33​}​]

反向传播

由这一篇《优雅地理解神经网络反向传播》可知，神经网络反向传播中，每一层计算损失函数关于该层输入的梯度，将其传给前一层。 (*)
假定输出层损失函数为 $L$

平均池化

$$\left[\begin{array}{ccc}\frac{\partial L}{\partial x_{11}}& \frac{\partial L}{\partial x_{12}}& \frac{\partial L}{\partial x_{13}}\\ & & \\ \frac{\partial L}{\partial x_{21}}& \frac{\partial L}{\partial x_{22}}& \frac{\partial L}{\partial x_{23}}\\ & & \\ \frac{\partial L}{\partial x_{31}}& \frac{\partial L}{\partial x_{32}}& \frac{\partial L}{\partial x_{33}}\end{array}\right]\leftarrow \left[\begin{array}{cc}\frac{\partial L}{\partial z_{11}}& \frac{\partial L}{\partial z_{12}}\\ & \\ \frac{\partial L}{\partial z_{21}}& \frac{\partial L}{\partial z_{22}}\end{array}\right]$$
也就是按照箭头方向，已知所有 $\frac{\partial L}{\partial z_{ij}}$ 求解所有 $\frac{\partial L}{\partial x_{ij}}$ 的过程。

比如上图中 $(i,j)=(2,2)$ 的情形， $x_{22}$ 对最终损失函数 $L$ 的贡献反映在四项 $\frac{\partial L}{\partial z_{11}},\frac{\partial L}{\partial z_{12}},\frac{\partial L}{\partial z_{21}},\frac{\partial L}{\partial z_{22}}$ 上。

于是由链式法则，

$$\begin{gathered} \frac{\partial L}{\partial x_{22}}=\frac{\partial L}{\partial z_{11}}\frac{\partial z_{11}}{\partial x_{22}}+\frac{\partial L}{\partial z_{12}}\frac{\partial z_{12}}{\partial x_{22}}+\frac{\partial L}{\partial z_{21}}\frac{\partial z_{21}}{\partial x_{22}}+\frac{\partial L}{\partial z_{22}}\frac{\partial z_{22}}{\partial x_{22}} \\ =\frac{1}{4}(\frac{\partial L}{\partial z_{11}}+\frac{\partial L}{\partial z_{12}}+\frac{\partial L}{\partial z_{21}}+\frac{\partial L}{\partial z_{22}}) \end{gathered}$$

同理，易求得

$$\begin{gathered} \left[\begin{array}{ccc}\frac{\partial L}{\partial x_{11}}& \frac{\partial L}{\partial x_{12}}& \frac{\partial L}{\partial x_{13}}\\ & & \\ \frac{\partial L}{\partial x_{21}}& \frac{\partial L}{\partial x_{22}}& \frac{\partial L}{\partial x_{23}}\\ & & \\ \frac{\partial L}{\partial x_{31}}& \frac{\partial L}{\partial x_{32}}& \frac{\partial L}{\partial x_{33}}\end{array}\right] \\ =\frac{1}{4}\left[\begin{array}{ccc}\frac{\partial L}{\partial z_{11}}& \frac{\partial L}{\partial z_{11}}+\frac{\partial L}{\partial z_{12}}& \frac{\partial L}{\partial z_{12}}\\ & & \\ \frac{\partial L}{\partial x_{11}}+\frac{\partial L}{\partial z_{21}}& \frac{\partial L}{\partial z_{11}}+\frac{\partial L}{\partial z_{12}}+\frac{\partial L}{\partial z_{21}}+\frac{\partial L}{\partial z_{22}}& \frac{\partial L}{\partial x_{12}}+\frac{\partial L}{\partial z_{22}}\\ & & \\ \frac{\partial L}{\partial z_{21}}& \frac{\partial L}{\partial z_{21}}+\frac{\partial L}{\partial z_{22}}& \frac{\partial L}{\partial z_{22}}\end{array}\right] \end{gathered}$$
为了方便边界处理，可以填充（padding），也可以写成如下形式：

$$=\frac{1}{4}(\frac{\partial L}{\partial z_{11}}\left[\begin{array}{ccc}1& 1& 0\\ & & \\ 1& 1& 0\\ & & \\ 0& 0& 0\end{array}\right]+\frac{\partial L}{\partial z_{12}}\left[\begin{array}{ccc}0& 1& 1\\ & & \\ 0& 1& 1\\ & & \\ 0& 0& 0\end{array}\right]+\frac{\partial L}{\partial z_{21}}\left[\begin{array}{ccc}0& 0& 0\\ & & \\ 1& 1& 0\\ & & \\ 1& 1& 0\end{array}\right]+\frac{\partial L}{\partial z_{22}}\left[\begin{array}{ccc}0& 0& 0\\ & & \\ 0& 1& 1\\ & & \\ 0& 1& 1\end{array}\right])$$

可以理解为：
对于每个 $z_{ij}$, 将其对损失函数的贡献（偏微分）分配给求得它的 $x_{i^{\prime }{j}^{\prime }}$。

最大池化

池化核在每一个位置处时，只有该区域中最大的 $x$ 对损失函数有贡献。其表现类似于 relu 激活函数。需要记录最大的 $x$ 所在的位置。
例如假设前向传播时
$$\left[\begin{array}{ccc}{x}_{11}& x_{12}& x_{13}\\ x_{21}& x_{22}& x_{23}\\ x_{31}& x_{32}& x_{33}\end{array}\right]\to \left[\begin{array}{cc}{x}_{(i,j{)}_{1,1}}& x_{(i,j{)}_{1,2}}\\ x_{(i,j{)}_{2,1}}& x_{(i,j{)}_{2,2}}\end{array}\right]$$
则反向传播时

$$\left[\begin{array}{ccc}\frac{\partial L}{\partial x_{11}}& \frac{\partial L}{\partial x_{12}}& \frac{\partial L}{\partial x_{13}}\\ & & \\ \frac{\partial L}{\partial x_{21}}& \frac{\partial L}{\partial x_{22}}& \frac{\partial L}{\partial x_{23}}\\ & & \\ \frac{\partial L}{\partial x_{31}}& \frac{\partial L}{\partial x_{32}}& \frac{\partial L}{\partial x_{33}}\end{array}\right]=\frac{\partial L}{\partial z_{11}}{\delta }_{(i,j{)}_{1,1}}+\frac{\partial L}{\partial z_{12}}{\delta }_{(i,j{)}_{1,2}}+\frac{\partial L}{\partial z_{21}}{\delta }_{(i,j{)}_{2,1}}+\frac{\partial L}{\partial z_{22}}{\delta }_{(i,j{)}_{2,2}}$$

其中 ${\delta }_{(i,j)}$ 表示只在 $(i,j)$ 处为1，其余地方为0的矩阵。