机器学习 (三): 反向传播数学推导 & 神经网络

超、凢脫俗 2022-03-15 05:26 295阅读 0赞

### 反向传播推导 & 神经网络 ###

*  约定
 *  推导
 *  神经网络实现

# 约定 #

本小节所使用的神经网络如图所示  
![神经网络][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ZpbmNlZWVfXw_size_16_color_FFFFFF_t_70]  
其中

*  一共有三层, 序号分别是 0, 1, 2 (第0层为输入层)
 *  输入为向量  x x x 有两个项  x 1 x\_1 x1 和  x 2 x\_2 x2
 *   a ( i ) a^\{(i)\} a(i) 表示第 i 层的输出向量,  a j ( i ) a\_j^\{(i)\} aj(i) 表示第 i 层输出向量的第 j 项. 注意  x = a ( 0 ) x = a^\{(0)\} x=a(0)
 *   W ( i ) W^\{(i)\} W(i) 表示连接第 i-1 层和第 i 层的边, 即系数,  w j k ( i ) w\_\{jk\}^\{(i)\} wjk(i) 表示连接第 i 层第 j 个节点和第 i-1 层第 k 个节点的边, 注意 jk 的顺序是从右到左连接
 *   z ( i ) = W ( i ) × a ( i − 1 ) z^\{(i)\} = W^\{(i)\} \\times a^\{(i-1)\} z(i)=W(i)×a(i−1)
 *   a ( i ) = s i g m o i d ( z ( i ) ) a^\{(i)\} = sigmoid(z^\{(i)\}) a(i)=sigmoid(z(i)) 也就是说本网络使用的激活函数是 sigmoid
 *   s i g m o i d ( x ) = 1 1 + e − x sigmoid(x) = \\frac\{1\}\{1 + e^\{-x\}\} sigmoid(x)=1\+e−x1
 *  损失函数为  J ( W , x , y ) = ∑ i = 1 2 \[ y l n ( a i ( 2 ) ) + ( 1 − y ) l n ( 1 − a i ( 2 ) ) \] J(W, x, y) = \\sum\_\{i=1\}^2\[yln(a\_i^\{(2)\}) + (1-y)ln(1 - a\_i^\{(2)\})\] J(W,x,y)=∑i=12\[yln(ai(2))\+(1−y)ln(1−ai(2))\]
 *  简便起见不考虑 bias 项

# 推导 #

我们的目标是最小化  J ( W , x , y ) J(W, x, y) J(W,x,y), 其中 W 是所有的边, 而 x, y 是测试数据集为常量, 所以我们需要求  δ J δ W \\frac\{\\delta J\}\{\\delta W\} δWδJ, 具体地, 我们要求  δ J δ w ( 2 ) \\frac\{\\delta J\}\{\\delta w^\{(2)\}\} δw(2)δJ 和  δ J δ w ( 1 ) \\frac\{\\delta J\}\{\\delta w^\{(1)\}\} δw(1)δJ :

δ J δ w ( 2 ) = δ J δ a ( 2 ) × δ a ( 2 ) δ z ( 2 ) × δ z ( 2 ) δ w ( 2 ) \\frac\{\\delta J\}\{\\delta w^\{(2)\}\} = \\frac\{\\delta J\}\{\\delta a^\{(2)\}\} \\times \\frac\{\\delta a^\{(2)\}\}\{\\delta z^\{(2)\}\} \\times \\frac\{\\delta z^\{(2)\}\}\{\\delta w^\{(2)\}\} δw(2)δJ=δa(2)δJ×δz(2)δa(2)×δw(2)δz(2)

δ J δ a ( 2 ) = ( y l n ( a ( 2 ) ) + ( 1 − y ) l n ( 1 − a ( 2 ) ) ) ′ = y a ( 2 ) − 1 − y 1 − a ( 2 ) \\frac\{\\delta J\}\{\\delta a^\{(2)\}\} = (yln(a^\{(2)\}) + (1-y)ln(1-a^\{(2)\}))^\{&\#x27;\} = \\frac\{y\}\{a^\{(2)\}\} - \\frac\{1-y\}\{1 - a^\{(2)\}\} δa(2)δJ=(yln(a(2))\+(1−y)ln(1−a(2)))′=a(2)y−1−a(2)1−y

δ a ( 2 ) δ z ( 2 ) = a ( 2 ) ⋅ ( 1 − a ( 2 ) ) \\frac\{\\delta a^\{(2)\}\}\{\\delta z^\{(2)\}\} = a^\{(2)\} \\cdot (1-a^\{(2)\}) δz(2)δa(2)=a(2)⋅(1−a(2))

由上述二式可得

δ J δ z ( 2 ) = δ J δ a ( 2 ) × δ a ( 2 ) δ z ( 2 ) = y ⋅ ( 1 − a ( 2 ) ) − ( 1 − y ) ⋅ a ( 2 ) \\frac\{\\delta J\}\{\\delta z^\{(2)\}\} = \\frac\{\\delta J\}\{\\delta a^\{(2)\}\} \\times \\frac\{\\delta a^\{(2)\}\}\{\\delta z^\{(2)\}\} = y \\cdot (1-a^\{(2)\}) - (1-y)\\cdot a^\{(2)\} δz(2)δJ=δa(2)δJ×δz(2)δa(2)=y⋅(1−a(2))−(1−y)⋅a(2)

因为 y 的值只可能是 0 或 1, 当 y = 0 时, 上式等于

− a ( 2 ) = 0 − a ( 2 ) = y − a ( 2 ) -a^\{(2)\} = 0-a^\{(2)\} = y - a^\{(2)\} −a(2)=0−a(2)=y−a(2)

当 y = 1 时, 上式等于

1 − a ( 2 ) = y − a ( 2 ) 1 - a^\{(2)\} = y - a^\{(2)\} 1−a(2)=y−a(2)

所以

δ J δ z ( 2 ) = a ( 2 ) − y \\frac\{\\delta J\}\{\\delta z^\{(2)\}\} = a^\{(2)\} - y δz(2)δJ=a(2)−y

令  δ ( 2 ) = δ J δ z ( 2 ) = a ( 2 ) − y \\delta^\{(2)\} = \\frac\{\\delta J\}\{\\delta z^\{(2)\}\} = a^\{(2)\} - y δ(2)=δz(2)δJ=a(2)−y

因为

z ( 2 ) = W ( 2 ) × a ( 1 ) z^\{(2)\} = W^\{(2)\} \\times a^\{(1)\} z(2)=W(2)×a(1)

所以

δ z ( 2 ) δ W ( 2 ) = a ( 1 ) \\frac\{\\delta z^\{(2)\}\}\{\\delta W^\{(2)\}\} = a^\{(1)\} δW(2)δz(2)=a(1)

综上

δ J δ w ( 2 ) = δ J δ a ( 2 ) × δ a ( 2 ) δ z ( 2 ) × δ z ( 2 ) δ W ( 2 ) = δ ( 2 ) × a ( 1 ) \\frac\{\\delta J\}\{\\delta w^\{(2)\}\} = \\frac\{\\delta J\}\{\\delta a^\{(2)\}\} \\times \\frac\{\\delta a^\{(2)\}\}\{\\delta z^\{(2)\}\} \\times \\frac\{\\delta z^\{(2)\}\}\{\\delta W^\{(2)\}\} = \\delta^\{(2)\} \\times a^\{(1)\} δw(2)δJ=δa(2)δJ×δz(2)δa(2)×δW(2)δz(2)=δ(2)×a(1)

--------------------

现在开始求  δ J δ w ( 1 ) \\frac\{\\delta J\}\{\\delta w^\{(1)\}\} δw(1)δJ:

δ J δ w ( 1 ) = δ J δ a ( 1 ) × δ a ( 1 ) δ z ( 1 ) × δ z ( 1 ) δ W ( 1 ) \\frac\{\\delta J\}\{\\delta w^\{(1)\}\} = \\frac\{\\delta J\}\{\\delta a^\{(1)\}\} \\times \\frac\{\\delta a^\{(1)\}\}\{\\delta z^\{(1)\}\} \\times \\frac\{\\delta z^\{(1)\}\}\{\\delta W^\{(1)\}\} δw(1)δJ=δa(1)δJ×δz(1)δa(1)×δW(1)δz(1)

因为

z ( 2 ) = W ( 2 ) × a ( 1 ) z^\{(2)\} = W^\{(2)\} \\times a^\{(1)\} z(2)=W(2)×a(1)

所以

δ z ( 2 ) δ a ( 1 ) = W ( 2 ) \\frac\{\\delta z^\{(2)\}\}\{\\delta a^\{(1)\}\} = W^\{(2)\} δa(1)δz(2)=W(2)

所以

δ J δ a ( 1 ) = δ J δ z ( 2 ) × δ z ( 2 ) δ a ( 1 ) = W ( 2 ) T × δ ( 2 ) \\frac\{\\delta J\}\{\\delta a^\{(1)\}\} = \\frac\{\\delta J\}\{\\delta z^\{(2)\}\} \\times \\frac\{\\delta z^\{(2)\}\}\{\\delta a^\{(1)\}\} =W^\{(2)T\} \\times \\delta^\{(2)\} δa(1)δJ=δz(2)δJ×δa(1)δz(2)=W(2)T×δ(2)

因为

δ a ( 1 ) δ z ( 1 ) = a ( 1 ) ⋅ ( 1 − a ( 1 ) ) \\frac\{\\delta a^\{(1)\}\}\{\\delta z^\{(1)\}\} = a^\{(1)\} \\cdot(1-a^\{(1)\}) δz(1)δa(1)=a(1)⋅(1−a(1))  
所以

δ J δ z ( 1 ) = δ J δ a ( 1 ) × δ a ( 1 ) δ z ( 1 ) = W ( 2 ) T × δ ( 2 ) ⋅ a ( 1 ) ⋅ ( 1 − a ( 1 ) ) \\frac\{\\delta J\}\{\\delta z^\{(1)\}\} = \\frac\{\\delta J\}\{\\delta a^\{(1)\}\} \\times \\frac\{\\delta a^\{(1)\}\}\{\\delta z^\{(1)\}\} = W^\{(2)T\} \\times \\delta^\{(2)\} \\cdot a^\{(1)\} \\cdot (1 - a^\{(1)\}) δz(1)δJ=δa(1)δJ×δz(1)δa(1)=W(2)T×δ(2)⋅a(1)⋅(1−a(1))

令  δ ( 1 ) = δ J δ z ( 1 ) = W ( 2 ) T × δ ( 2 ) ⋅ a ( 1 ) ⋅ ( 1 − a ( 1 ) ) \\delta^\{(1)\} = \\frac\{\\delta J\}\{\\delta z^\{(1)\}\} = W^\{(2)T\} \\times \\delta^\{(2)\} \\cdot a^\{(1)\} \\cdot (1 - a^\{(1)\}) δ(1)=δz(1)δJ=W(2)T×δ(2)⋅a(1)⋅(1−a(1))

又因为

z ( 1 ) = W ( 1 ) × a ( 0 ) z^\{(1)\} = W^\{(1)\} \\times a^\{(0)\} z(1)=W(1)×a(0)

所以

δ z ( 1 ) δ W ( 1 ) = a ( 0 ) \\frac\{\\delta z^\{(1)\}\}\{\\delta W^\{(1)\}\} = a^\{(0)\} δW(1)δz(1)=a(0)

综上

δ J δ w ( 1 ) = δ J δ a ( 1 ) × δ a ( 1 ) δ z ( 1 ) × δ z ( 1 ) δ W ( 1 ) = δ ( 1 ) × a ( 0 ) \\frac\{\\delta J\}\{\\delta w^\{(1)\}\} = \\frac\{\\delta J\}\{\\delta a^\{(1)\}\} \\times \\frac\{\\delta a^\{(1)\}\}\{\\delta z^\{(1)\}\} \\times \\frac\{\\delta z^\{(1)\}\}\{\\delta W^\{(1)\}\} = \\delta^\{(1)\} \\times a^\{(0)\} δw(1)δJ=δa(1)δJ×δz(1)δa(1)×δW(1)δz(1)=δ(1)×a(0)

# 神经网络实现 #

import numpy as np
    import matplotlib.pyplot as plt 
    
    def init_model(X, Y, layers):
    	model = { }
    	nodes = []
    	nodes.append(X.shape[0])
    	for l in layers:
    		nodes.append(l)
    	nodes.append(Y.shape[0])
    	model['depth'] = len(layers) + 1
    	for n in range(model['depth']):
    		model['W' + str(n + 1)] = np.random.rand(nodes[n+1], nodes[n])
    		model['B' + str(n + 1)] = np.random.rand(nodes[n+1], 1)
    	return model
    
    def sigmoid(x):
    	return 1.0 / (1 + np.exp(-x))
    
    def forward_propagation(X, Y, model):
    	result = X.copy()
    	model['A0'] = result.copy()
    	for i in range(model['depth']):
    		W = model['W' + str(i+1)]
    		B = model['B' + str(i+1)]
    		result = np.dot(W, result) + B
    		model['Z' + str(i+1)] = result.copy()
    		result = sigmoid(result)
    		model['A' + str(i+1)] = result.copy()
    	return result
    
    def cross_entropy(Y, result):
    	return -Y * np.log(result) - (1 - Y) * np.log(1 - result)
    
    def cost(Y, result):
    	return np.sum(cross_entropy(Y, result)) / Y.shape[1]
    
    def backward_propagation(Y, model, learning_rate):
    	L = model['depth']
    	m = Y.shape[1]
    	delta = model['A' + str(L)] - Y
    	model['dW' + str(L)] = np.dot(delta, model['A' + str(L-1)].T) / m
    	model['dB' + str(L)] = np.sum(delta, axis = 1, keepdims=True) / m
    	for l in range(L - 1, 0, -1):
    		delta = (np.dot(model['W' + str(l+1)].T, delta) * model['A' + str(l)] * (1 - model['A' + str(l)])) 
    		model['dW' + str(l)] = np.dot(delta, model['A' + str(l-1)].T) / m
    		model['dB' + str(l)] = np.sum(delta, axis=1, keepdims=True) / m
    
    	for l in range(model['depth'], 0, -1):
    		model['W' + str(l)] -= learning_rate * model['dW' + str(l)]
    		model['B' + str(l)] -= learning_rate * model['dB' + str(l)]
    
    def train(X, Y, model, learning_rate = 0.05, times = 50, show_cost_history = True):
    	cost_history = []
    	result = 0
    	for t in range(times):
    		result = forward_propagation(X, Y, model)
    		backward_propagation(Y, model, learning_rate)
    		cost_history.append(cost(Y, result))
    	if show_cost_history:
    		plt.plot(cost_history)
    		plt.show()
    
    if __name__ == '__main__':
    	m = 1000
    	X = np.random.rand(2, m)
    	Y = np.zeros((1, m))
    	for i in range(m):
    		if (X[0, i] < 0.5) and (X[1, i] < 0.5):
    			Y[0, i] = 1
    
    	model = init_model(X, Y, [2])
    	train(X, Y, model)

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ZpbmNlZWVfXw_size_16_color_FFFFFF_t_70]: /images/20220315/19b7d47e0d9c4a31acdeea32b91a0645.png