Coursera机器学习笔记 第5周 第九章 神经网络 相关数学公式证明

神经网络 相关数学公式证明

0. Coursera机器学习神经网络的其他笔记：

1. Coursera机器学习笔记 第4周 第八章 神经网络：表述（一）
2. Coursera机器学习笔记 第4周 第八章 神经网络：表述（二）
3. Coursera机器学习笔记 第5周 第九章 神经网络：学习（一）
4. Coursera机器学习笔记 第5周 第九章 神经网络：学习（二）

1. 神经网络的回顾

我们仍然沿用上图中的神经网络结构。该网络共有四层，输入层有3个神经元，第二、三层均有5个神经元，输出层有4个神经元（偏置单元未画出），所以权重矩阵$\Theta^{(1)}, \Theta^{(2)}, \Theta^{(3)}$的大小分别是$[5 \times 4], [5 \times 6], [4 \times 6]$。权重矩阵如下：

$$\Theta^{(1)}=\begin{bmatrix} \Theta^{(1)}_{10} & ... & \Theta^{(1)}_{13}\\ ... & ... & ...\\ \Theta^{(1)}_{50} & ... & \Theta^{(1)}_{53} \end{bmatrix} \Theta^{(2)}=\begin{bmatrix} \Theta^{(2)}_{10} & ... & \Theta^{(2)}_{15}\\ ... & ... & ...\\ \Theta^{(2)}_{50} & ... & \Theta^{(2)}_{55} \end{bmatrix} \Theta^{(3)}=\begin{bmatrix} \Theta^{(3)}_{10} & ... & \Theta^{(1)}_{15}\\ ... & ... & ...\\ \Theta^{(3)}_{40} & ... & \Theta^{(1)}_{45} \end{bmatrix}$$
为了和代价函数的形式统一，我们用 j 和 i 分别表示权重矩阵的行和列。注意：j=1 代表矩阵的第 1 行，i=1 代表矩阵的第 2 列（第 1 列相当于线性回归中的$\theta_0$）。

神经网络的一般化代价函数为：

$$\begin{gather*}J(\Theta) = - \frac{1}{m} \sum_{t=1}^m\sum_{k=1}^K \left[ y^{(t)}_k \ \log (h_\Theta (x^{(t)})_k) + (1 - y^{(t)}_k)\ \log (1 - h_\Theta(x^{(t)})_k)\right] + \frac{\lambda}{2m}\sum_{l=1}^{L-1} \sum_{i=1}^{s_l} \sum_{j=1}^{s_l+1} ( \Theta_{j,i}^{(l)})^2\end{gather*}$$

我们只用一条数据(x,y)，并且忽略正则化，则代价函数(K=4)为：

$$\begin{gather*}Cost(x) = \sum_{k=1}^K Cost(x)_k \end{gather*} \\ Cost(x)_k = - \left[ y_k \ \log (h_\Theta (x)_k) + (1 - y_k)\ \log (1 - h_\Theta(x)_k)\right]$$

2. 证明$\dfrac{\partial Cost(x)}{\partial \Theta^{(3)}}$

针对该神经网络，声明一些事实：

• $h_\Theta (x)$就是$a^{(4)}$，即$h_\Theta (x)=a^{(4)}$
• $a^{(4)} = g(z^{(4)})$，且$\dfrac{\partial a^{(4)}}{\partial z^{(4)}} = a^{(4)} (1 - a^{(4)})$
• $z^{(4)}_j = \sum^{5}_{i=0}\Theta^{(3)}_{ji}a^{(3)}_i$，其中 $1 \leqslant j \leqslant 4$
• j=1 代表矩阵的第 1 行，i=1 代表矩阵的第 2 列（第 1 列相当于线性回归中的$\theta_0$）

举例：

$$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{10}^{(3)}} &= \dfrac{\partial \sum^K_{k=1}Cost(x)_k}{\partial \Theta_{10}^{(3)}} \\ &= \sum^K_{k=1} \dfrac{\partial Cost(x)_k}{\partial \Theta_{10}^{(3)}} \\ &= \dfrac{\partial Cost(x)_1}{\partial \Theta_{10}^{(3)}} + 0 + 0 + 0 \\ &= \dfrac{\partial Cost(x)_1}{\partial h_{\Theta}(x)_1} \times \dfrac{\partial h_{\Theta}(x)_1}{\partial z^{(4)}_1} \times \dfrac{\partial z^{(4)}_1}{\partial \Theta_{10}^{(3)}} \\ &= -[y_1\dfrac{1}{h_{\Theta}(x)_1}+(1-y_1)\dfrac{-1}{1-h_{\Theta}(x)_1}] \times h_{\Theta}(x)_1(1-h_{\Theta}(x)_1) \times a^{(3)}_0 \\ &= [h_{\Theta}(x)_1 - y_1]a^{(3)}_0 \\ &= [a^{(4)}_1 - y_1]a^{(3)}_0 \end{align*}$$ $$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{20}^{(3)}} &= 0 + \dfrac{\partial Cost(x)_2}{\partial \Theta_{20}^{(3)}} + 0 + 0 \\ &= \dfrac{\partial Cost(x)_2}{\partial h_{\Theta}(x)_2} \times \dfrac{\partial h_{\Theta}(x)_2}{\partial z^{(4)}_2} \times \dfrac{\partial z^{(4)}_2}{\partial \Theta_{20}^{(3)}} \\ &= -[y_2\dfrac{1}{h_{\Theta}(x)_2}+(1-y_2)\dfrac{-1}{1-h_{\Theta}(x)_2}] \times h_{\Theta}(x)_2(1-h_{\Theta}(x)_2) \times a^{(3)}_0 \\ &= [h_{\Theta}(x)_2 - y_2]a^{(3)}_0 \end{align*}$$ $$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{21}^{(3)}} &= 0 + \dfrac{\partial Cost(x)_2}{\partial \Theta_{21}^{(3)}} + 0 + 0 \\ &= \dfrac{\partial Cost(x)_2}{\partial h_{\Theta}(x)_2} \times \dfrac{\partial h_{\Theta}(x)_2}{\partial z^{(4)}_2} \times \dfrac{\partial z^{(4)}_2}{\partial \Theta_{21}^{(3)}} \\ &= -[y_2\dfrac{1}{h_{\Theta}(x)_2}+(1-y_2)\dfrac{-1}{1-h_{\Theta}(x)_2}] \times h_{\Theta}(x)_2(1-h_{\Theta}(x)_2) \times a^{(3)}_1 \\ &= [h_{\Theta}(x)_2 - y_2]a^{(3)}_1 \end{align*}$$
综上： $$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{ji}^{(3)}} &= \dfrac{\partial Cost(x)_j}{\partial h_{\Theta}(x)_j} \times \dfrac{\partial h_{\Theta}(x)_j}{\partial z^{(4)}_j} \times \dfrac{\partial z^{(4)}_j}{\partial \Theta_{ji}^{(3)}} \\ &= -[y_j\dfrac{1}{h_{\Theta}(x)_j}+(1-y_j)\dfrac{-1}{1-h_{\Theta}(x)_j}] \times h_{\Theta}(x)_j(1-h_{\Theta}(x)_j) \times a^{(3)}_i \\ &= [h_{\Theta}(x)_j - y_j]（a^{(3)}_i \end{align*}$$
引入 $\color{Red}{\delta^{(4)} = a^{(4)} - y}$，再加上 $h_\Theta (x) =a^{(4)}$ ，所以 $\dfrac{\partial Cost(x)}{\partial \Theta_{ji}^{(3)}} =\delta^{(4)}_j a^{(3)}_i$
把矩阵下标去掉，公式变为： $\dfrac{\partial Cost(x)}{\partial \Theta^{(3)}} =\color{Red}{\delta^{(4)}}(a^{(3)})^T$

3. 证明$\dfrac{\partial Cost(x)}{\partial \Theta^{(2)}}，\dfrac{\partial Cost(x)}{\partial \Theta^{(1)}}$

举例：

$$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{10}^{(2)}} &= \sum^{K}_{k=1}\dfrac{\partial Cost(x)_k}{\partial \Theta_{10}^{(2)}} \\ &= \sum^{K}_{k=1} \left[\dfrac{\partial Cost(x)_k}{\partial h_{\Theta}(x)_k} \times \dfrac{\partial h_{\Theta}(x)_k}{\partial z^{(4)}_k} \times \dfrac{\partial z^{(4)}_k}{\partial a^{(3)}_1} \times \dfrac{\partial a^{(3)}_1}{\partial z^{(3)}_1} \times \dfrac{\partial z^{(3)}_1}{\partial \Theta^{(2)}_{10}}\right] \\ &= \sum^{K}_{k=1} \left[(h_{\Theta}(x)_k-y_k) \times \Theta^{(3)}_{k1} \times a^{(3)}_1(1-a^{(3)}_1) \times a^{(2)}_0\right]\end{align*}$$ $$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{20}^{(2)}} &= \sum^{K}_{k=1}\dfrac{\partial Cost(x)_k}{\partial \Theta_{20}^{(2)}} \\ &= \sum^{K}_{k=1} \left[\dfrac{\partial Cost(x)_k}{\partial h_{\Theta}(x)_k} \times \dfrac{\partial h_{\Theta}(x)_k}{\partial z^{(4)}_k} \times \dfrac{\partial z^{(4)}_k}{\partial a^{(3)}_2} \times \dfrac{\partial a^{(3)}_2}{\partial z^{(3)}_2} \times \dfrac{\partial z^{(3)}_2}{\partial \Theta^{(2)}_{20}}\right] \\ &= \sum^{K}_{k=1} \left[(h_{\Theta}(x)_k-y_k) \times \Theta^{(3)}_{k2} \times a^{(3)}_2(1-a^{(3)}_2) \times a^{(2)}_0\right]\end{align*}$$ $$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{21}^{(2)}} &= \sum^{K}_{k=1}\dfrac{\partial Cost(x)_k}{\partial \Theta_{21}^{(2)}} \\ &= \sum^{K}_{k=1} \left[\dfrac{\partial Cost(x)_k}{\partial h_{\Theta}(x)_k} \times \dfrac{\partial h_{\Theta}(x)_k}{\partial z^{(4)}_k} \times \dfrac{\partial z^{(4)}_k}{\partial a^{(3)}_2} \times \dfrac{\partial a^{(3)}_2}{\partial z^{(3)}_2} \times \dfrac{\partial z^{(3)}_2}{\partial \Theta^{(2)}_{21}}\right] \\ &= \sum^{K}_{k=1} \left[(h_{\Theta}(x)_k-y_k) \times \Theta^{(3)}_{k2} \times a^{(3)}_2(1-a^{(3)}_2) \times a^{(2)}_1\right]\end{align*}$$ $$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{32}^{(2)}} &= \sum^{K}_{k=1}\dfrac{\partial Cost(x)_k}{\partial \Theta_{32}^{(2)}} \\ &= \sum^{K}_{k=1} \left[\dfrac{\partial Cost(x)_k}{\partial h_{\Theta}(x)_k} \times \dfrac{\partial h_{\Theta}(x)_k}{\partial z^{(4)}_k} \times \dfrac{\partial z^{(4)}_k}{\partial a^{(3)}_3} \times \dfrac{\partial a^{(3)}_3}{\partial z^{(3)}_3} \times \dfrac{\partial z^{(3)}_3}{\partial \Theta^{(2)}_{32}}\right] \\ &= \sum^{K}_{k=1} \left[(h_{\Theta}(x)_k-y_k) \times \Theta^{(3)}_{k3} \times a^{(3)}_3(1-a^{(3)}_3) \times a^{(2)}_2\right]\end{align*}$$

记住$\color{Red}{\delta^{(4)} = a^{(4)} - y, h_\Theta (x) =a^{(4)} }$。根据前面的例子把公式进行一般化（其中$(\Theta^{(3)})^T_{j:}$表示矩阵$((\Theta^{(3)})^T$第 j 行，第一个乘号是矩阵相乘，其他乘号是实数相乘）：

$$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta_{ji}^{(2)}} &= \sum^{K}_{k=1} \left[(h_{\Theta}(x)_k-y_k) \times \Theta^{(3)}_{kj} \times a^{(3)}_j(1-a^{(3)}_j) \times a^{(2)}_i\right] \\ &=\sum^{K}_{k=1} \left[\delta^{(4)}_k \times \Theta^{(3)}_{kj}\right] \times a^{(3)}_j(1-a^{(3)}_j) \times a^{(2)}_i \\ &=(\Theta^{(3)})^T_{j:} \times \delta^{(4)} \times a^{(3)}_j(1-a^{(3)}_j) \times a^{(2)}_i\end{align*}$$

把矩阵下标去掉，公式变为（两个乘号表示矩阵相乘，.* 表示对应元素相乘）：

$$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta^{(2)}} &=((\Theta^{(3)})^T \times \delta^{(4)}) .*\ a^{(3)} .*\ (1-a^{(3)}) \times a^{(2)}\end{align*}$$

再引入$\color{Red}{\delta^{(3)} = ((\Theta^{(3)})^T \delta^{(4)})\ .*\ a^{(3)}\ .*\ (1 - a^{(3)}) }$，公式变为：

$$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta^{(2)}} &=\color{Red}{\delta^{(3)}} a^{(2)}\end{align*}$$

再引入$\color{Red}{\delta^{(2)} = ((\Theta^{(2)})^T \delta^{(3)})\ .*\ a^{(2)}\ .*\ (1 - a^{(2)}) }$，所以

$$\begin{align*}\dfrac{\partial Cost(x)}{\partial \Theta^{(1)}} &=\color{Red}{\delta^{(2)}} a^{(1)}\end{align*}$$

4. 总结

一条数据的反向传播如下：

1. 计算输出层的误差：$\delta^{(4)} = a^{(4)} - y$
2. 计算第三层的误差：$\delta^{(3)} = ((\Theta^{(3)})^T \delta^{(4)})\ .*\ a^{(3)}\ .*\ (1 - a^{(3)})$
3. 计算第二层的误差：$\delta^{(2)} = ((\Theta^{(2)})^T \delta^{(3)})\ .*\ a^{(2)}\ .*\ (1 - a^{(2)})$。第一层是输入变量，不存在误差，所以到第二层即可。
4. 为什么计算每一层的误差 $\delta$ ？因为经过一系列复杂的求导后，我们通过 $\delta$ 可以计算代价函数对每一层权重矩阵的每一个参数的偏导数（无正则化处理或$\lambda=0$）：$\dfrac{\partial}{\partial \Theta_{i,j}^{(l)}}J(\Theta) = \delta^{(l+1)}_i a^{(l)}_j$，其中每个元素都是一个实数！

重点：

• 误差公式：$\color{Red}{\delta^{(l)} = ((\Theta^{(l)})^T \delta^{(l+1)})\ .*\ g'(z^{(l)})}$，其中$\color{Red}{g'(z^{(l)}) = a^{(l)}\ .*\ (1 - a^{(l)})}$
• 偏导数（梯度）公式：$\dfrac{\partial}{\partial \Theta_{i,j}^{(l)}}J(\Theta) = \delta^{(l+1)}_i a^{(l)}_j$
• 偏导数（梯度）公式（矩阵形式）：$\dfrac{\partial}{\partial \Theta^{(l)}}J(\Theta) = \delta^{(l+1)} (a^{(l)})^T$