backpropagation算法示例

backpropagation算法示例

下面举个例子，假设在某个mini-batch的有样本X和标签Y，其中\(X\in R^{m\times 2}, Y\in R^{m\times 1}\)，现在有个两层的网络，对应的计算如下：
\[
\begin{split}
i_1 &= XW_1+ b_1\\
o_1 &= sigmoid(i_1)\\
i_2 &= o_1W_2 + b_2\\
o_2 &= sigmoid(i_2)
\end{split}
\]
其中\(W_1 \in R^{2\times 3}, b_1\in R^{1\times 3}, W_2\in R^{3\times 1}, b_2\in R^{1\times 1}\)都是参数，然后使用平方损失函数
\[
cost = \dfrac{1}{2m}\sum_i^m(o_{2i} - Y_i)^2
\]
下面给出反向传播的过程
\[
\begin{split}
\dfrac{\partial cost}{\partial o_2} &= \dfrac{1}{m}(o_2 - Y)\\
\dfrac{\partial o_2}{\partial i_2} &= sigmoid(i_2)\odot (1 - sigmoid(i_2)) = o_2 \odot (1 - o_2)\\
\dfrac{\partial i_2}{\partial W_2} &= o_1\\
\dfrac{\partial i_2}{\partial o_2} &= w_2\\
\dfrac{\partial i_2}{\partial b_2} &= 1\\
\dfrac{\partial o_1}{\partial i_1} &= sigmoid(i_1)\odot (1 - sigmoid(i_1)) = o_1\odot (1 - o_1)\\
\dfrac{\partial i_1}{\partial W_1} &= X\\
\dfrac{\partial i_1}{\partial b_1} &= 1
\end{split}
\]
所以有

\[
\begin{split}
\Delta W_2 &= \dfrac{\partial cost}{\partial o_2}\dfrac{\partial o_2}{\partial i_2}\dfrac{\partial i_2}{\partial W_2}\\
\Delta b_2 &= \dfrac{\partial cost}{\partial o_2}\dfrac{\partial o_2}{\partial i_2}\dfrac{\partial i_2}{\partial b_2}\\
\Delta W_1 &= \dfrac{\partial cost}{\partial o_2}\dfrac{\partial o_2}{\partial i_2}\dfrac{\partial i_2}{\partial o_1}\dfrac{\partial o_1}{\partial i_1}\dfrac{\partial i_1}{\partial W_1}\\
\Delta b_1 &= \dfrac{\partial cost}{\partial o_2}\dfrac{\partial o_2}{\partial i_2}\dfrac{\partial i_2}{\partial o_1}\dfrac{\partial o_1}{\partial i_1}\dfrac{\partial i_1}{\partial b_1}
\end{split}
\]
根据上述的链式法则，有
\[
\begin{split}
\Delta W_2 &= \left((\dfrac{1}{m}(o_2 - Y)\odot(o_2\odot (1 - o_2)))^T\times o_1\right)^T\\
\Delta W_1 &= \left((((\dfrac{1}{m}(o_2 - Y)\odot (o_2\odot (1 - o_2)))\times W_2^T)\odot o_1\odot (1 - o_1))^T\times X\right)^T
\end{split}
\]