神经网络的数学推导及简单实现

2019-04-09 本文已影响0人迎风漂扬

如图所示，先画一个简单的神经网络，输入是一个向量（x1,x2,x3,x4）
对于每一个所有的隐藏层和输出层的节点都有：
$u_{j}=\sum_{i} w_{i j} x_{i}$
$x_{j}=f\left(u_{j}\right)$
$w_{i j}$ ：不同层之间节点的连接权重
$x_{i}$ ：上一层节点的输出值
$u_{j}$ ：该节点的输入值
$x_{j}$ ：该节点的输出值
$f$ ：激活函数，这里采用ReLU函数

self.activative = lambda x: np.maximum(0, x)        #激活函
self.dirivative = lambda x: 1 if x > 0 else 0       #导数

例如对所有隐藏层节点5，6，7：
$x_{5}=f\left(\sum_{i=1}^{4} w_{i 5} x_{i}\right)$
$x_{6}=f\left(\sum_{i=1}^{4} w_{i 6} x_{i}\right)$
$x_{7}=f\left(\sum_{i=1}^{4} w_{i 7} x_{i}\right)$

为了方便，可以进一步将其转换为矩阵形式：
$f\left(\left[ \begin{array}{cccc}{w_{15}} & {w_{25}} & {w_{35}} & {w_{45}} \\ {w_{16}} & {w_{26}} & {w_{36}} & {w_{46}} \\ {w_{17}} & {w_{27}} & {w_{37}} & {w_{47}}\end{array}\right] \left[ \begin{array}{c}{x_{1}} \\ {x_{2}} \\ {x_{3}} \\ {x_{4}}\end{array}\right]\right)=f\left(\left[ \begin{array}{c}{u_{5}} \\ {u_{6}} \\ {u_{7}}\end{array}\right]\right)=\left[ \begin{array}{c}{x_{5}} \\ {x_{6}} \\ {x_{7}}\end{array}\right]$

self.u_hidden = np.dot(self.W_hidden, self.input)
self.x_hidden = self.activative(self.u_hidden)

接下来是输出层的计算：
$f\left(\left[ \begin{array}{ccc}{w_{58}} & {w_{68}} & {w_{78}} \\ {w_{59}} & {w_{69}} & {w_{79}}\end{array}\right] \left[ \begin{array}{c}{x_{5}} \\ {x_{6}} \\ {x_{7}}\end{array}\right]\right)=f\left(\left[ \begin{array}{l}{u_{8}} \\ {u_{9}}\end{array}\right]\right)=\left[ \begin{array}{l}{x_{8}} \\ {x_{9}}\end{array}\right]=\left[ \begin{array}{l}{y_{1}} \\ {y_{2}}\end{array}\right]$

self.u_output = np.dot(self.W_ouput, self.x_hidden)
self.y = self.activative(self.u_output)

以上就是正向的计算过程，接下来讨论误差的反向传播过程，首先来看损失函数：
$E=\frac{1}{2} \sum_{i}\left(t_{i}-y_{i}\right)^{2}$ ， $E$ 是误差， $t_{i}$ 为训练数据的标记， $y_{i}$ 是输出

在训练的过程中，我们需要训练的也就是各层的权重，根据梯度下降算法：
$w_{j k}=w_{j k}-\eta \nabla w_{j k}$
$\eta$ ：学习率
$\nabla w_{j k}$ ： $w_{j k}$ 的梯度

首先来推导输出层的梯度，例如：
$\nabla w_{58}=\frac{\partial E}{\partial u_{58}}=\frac{\partial E}{\partial u_{8}} \frac{\partial u_{8}}{\partial w_{58}}$
因为变量 $w_{58}$ 是通过影响节点8的输入值 $u_{8}$ 进而再影响到误差 $E$ ，接下来一个一个的求解分量：
$\frac{\partial u_{8}}{\partial w_{58}}=\frac{\partial\left(x_{5} w_{58}+x_{6} w_{68}+x_{7} w_{78}\right)}{\partial w_{58}}=x_{5}$
$\frac{\partial E}{\partial u_{8}}=\frac{\partial E}{\partial x_{8}} \frac{\partial x_{8}}{\partial u_{8}}$
$\frac{\partial x_{8}}{\partial u_{8}}=f^{\prime}\left(u_{8}\right)$
$\frac{\partial E}{\partial x_{8}}=\frac{\partial E}{\partial y_{1}}=\frac{\partial \frac{1}{2} \sum_{i}\left(t_{i}-y_{i}\right)^{2}}{\partial y_{1}}=-\left(t_{1}-y_{1}\right)$

综合上式，得到以下结果：
$\nabla w_{58}=-\left(t_{1}-y_{1}\right) f^{\prime}\left(u_{8}\right) x_{5}$

将下标处理一下，得到公式：
$\nabla w_{j k}=-\left(t_{k}-y_{k}\right) f^{\prime}\left(u_{k}\right) x_{j}$ ，又设
$\delta_{k}=\left(t_{k}-y_{k}\right) f^{\prime}\left(u_{k}\right)$
$\delta_{k}$ 为该节点的误差项，所以最终的公式为：
$w_{j k}=w_{j k}+\eta \delta_{k} x_{j}$

用矩阵表示：
$\left[ \begin{array}{c}{\delta_{8}} \\ {\delta_{9}}\end{array}\right]=\left[ \begin{array}{l}{\left(t_{1}-y_{1}\right) f^{\prime}\left(u_{8}\right)} \\ {\left(t_{2}-y_{2}\right) f^{\prime}\left(u_{9}\right)}\end{array}\right]$

self.delta_output = (self.target - self.y) * self.dirivative(self.u_output)

$\left[ \begin{array}{ccc}{-\nabla w_{58}} & {-\nabla w_{68}} & {-\nabla w_{78}} \\ {-\nabla w_{59}} & {-\nabla w_{69}} & {-\nabla w_{79}}\end{array}\right]=\left[ \begin{array}{c}{\delta_{8}} \\ {\delta_{9}}\end{array}\right] \left[ \begin{array}{ccc}{x_{5}} & {x_{6}} & {x_{7}}\end{array}\right]$

self.nablaW_output = np.dot(self.delta_output, np.transpose(self.x_hidden))
self.W_ouput += self.learningrate * self.nablaW_output

以上，就是输出层的公式推导及简要的代码，对于隐藏层，例如误差 $E$ 关于 $w_{15}$ 的梯度的计算同样，只不过关于误差项有些区别：
$\nabla w_{15}=\frac{\partial E}{\partial w_{15}}=\frac{\partial E}{\partial u_{5}} \frac{\partial u_{5}}{\partial w_{15}}$
根据之前的讨论，同样另 $\frac{\partial E}{\partial u_{5}}=-\delta_{5}$ ，所以：
$\frac{\partial u_{5}}{\partial w_{15}}=x_{1}$
$-\frac{\partial E}{\partial u_{5}}=\delta_{5}=-\frac{\partial E}{\partial x_{5}} \frac{\partial x_{5}}{\partial u_{5}}$
$\frac{\partial x_{5}}{\partial u_{5}}=f^{\prime}\left(u_{5}\right)$

而 $\frac{\partial E}{\partial x_{5}}$ 的计算就稍微复杂一些，因为 $x_{5}$ 的值对其下层的节点（8，9）都有影响，所以根据全导数公式：
$\frac{\partial E}{\partial x_{5}}=\frac{\partial E}{\partial u_{8}} \frac{\partial u_{8}}{\partial x_{5}}+\frac{\partial E}{\partial u_{9}} \frac{\partial u_{9}}{\partial x_{5}}=\sum_{k=8}^{9} \frac{\partial E}{\partial u_{k}} \frac{\partial u_{k}}{\partial x_{5}}$

$\frac{\partial u_{k}}{\partial x_{5}}=\frac{\partial\left(\sum_{j} w_{j k} x_{j}\right)}{\partial x_{5}}=w_{5 k}$
$\frac{\partial E}{\partial x_{5}}==-\sum_{k=8}^{9} \delta_{k} w_{5 k}$

综合上述公式，得到：
$\delta_{5}=-\frac{\partial E}{\partial x_{5}} \frac{\partial x_{5}}{\partial u_{5}}=f^{\prime}\left(u_{5}\right) \sum_{k=8}^{9} \delta_{k} w_{5 k}$
$w_{15}=w_{15}-\eta \nabla w_{15}=w_{15}+\eta \delta_{5} x_{1}=w_{15}+f^{\prime}\left(u_{5}\right) x_{1} \sum_{k=8}^{9} \delta_{k} w_{5 k}$
整理一下，最终得到公式：
$\delta_{j}=f^{\prime}\left(u_{j}\right) \sum_{k} w_{j k} \delta_{k}$
$w_{i j}=w_{i j}+\eta \delta_{j} x_{i}$

详细过程如下：
$\left[ \begin{array}{ll}{\delta_{8}} & {\delta_{9}}\end{array}\right] \left[ \begin{array}{cc}{w_{58}} & {w_{68}} & {w_{78}} \\ {w_{59}} & {w_{69}} & {w_{79}}\end{array}\right]=\left[\sum_{k=8}^{9} \delta_{k} w_{5 k} \sum_{k=8}^{9} \delta_{k} w_{6 k} \sum_{k=8}^{9} \delta_{k} w_{7 k}\right]$

temp = np.dot(np.transpose(self.delta_output), self.W_ouput)    #临时变量

$\left[\sum_{k=8}^{9} \delta_{k} w_{5 k} \sum_{k=8}^{9} \delta_{k} w_{6 k} \sum_{k=8}^{9} \delta_{k} w_{7 k}\right] \times\left[f^{\prime}\left(u_{5}\right) \quad f^{\prime}\left(u_{6}\right) \quad f^{\prime}\left(u_{7}\right)\right]=\left[ \begin{array}{ccc}{\delta_{5}} & {\delta_{6}} & {\delta_{7}}\end{array}\right]$

self.delta_hidden = self.dirivative(self.u_hidden) * np.transpose(temp)

$\left[ \begin{array}{c}{\delta_{5}} \\ {\delta_{6}} \\ {\delta_{7}}\end{array}\right] \left[ \begin{array}{lll}{x_{1}} & {x_{2}} & {x_{3}} & {x_{4}}\end{array}\right]=\left[ \begin{array}{cccc}{-\nabla w_{15}} & {-\nabla w_{25}} & {-\nabla w_{35}} & {-\nabla w_{45}} \\ {-\nabla w_{16}} & {-\nabla w_{26}} & {-\nabla w_{36}} & {-\nabla w_{46}} \\ {-\nabla w_{17}} & {-\nabla w_{27}} & {-\nabla w_{37}} & {-\nabla w_{47}}\end{array}\right]$

self.nablaW_hidden = np.dot(self.delta_hidden, np.transpose(self.input))
self.W_hidden += self.learningrate * self.nablaW_hidden

至此，隐藏层的权重也更新完毕

神经网络的数学推导及简单实现

猜你喜欢

热点阅读