(图一)
(图一)中的神经网络模型是由多个感知器(perceptron)分几层组合而成,所谓感知器就是单层的神经网络(准确的说应该不叫神经网络咯),它只有一个输出节点,如(图二)所示:
(图二) 感知器
一个感知器就相当于一个线性分类器,而一层神经网络有多个隐藏节点的,就是多个感知器的组合,那么它其实就是多个线性分类器组合形成非线性分类器咯,如(图三)所示:
(图三)
一层感知器拟合能力就颇为强大,而多层的感知器组合起来,那拟合能力更没得说,可惜的是拟合能力虽然强大,但是求出准确拟合参数的算法不是太好,容易陷入局部最小,而且BP算法很“擅长”陷入局部最小,所谓局部最小,如(图四)所示,网络的权重被随机初始化后,然后求得梯度,然后用梯度更新参数,如果初始化的参数的点选择的不恰当,当梯度为0的点可能是一个使得代价J局部最小的点,而不是全局最小的,自然得到的网络权重也不是最好的。BP算法一直都有这样的问题,而且也容易因为网络规模大导致过拟合,好在最近深度学习提了一系列的trick改善了这些问题。比如用贪心预训练来改进初始化参数,相当于找到了一个好的初始点,严格的说是在正负阶段里主动修改了J的“地形”,这是个人的一些理解,最后再结合标签用传统的BP算法继续进行寻找全局最小,这个BP算法的作用在深度学习里也叫权重微调,当然BP不是唯一的权重微调算法,各种微调的宗旨只有一个:求取目标函数的梯度,更新参数。另外深度学习里利用稀疏和dropout来阻止过拟合。
(图四)
介绍了BP算法的作用,那么就来看下BP算法的原理,也很简单,就是把目标函数各层的权重进行求导,因为我们要更新权重就是要求出梯度,然后利用梯度更新权重。网络从输入到最终的输出经过了多层的函数处理,求导的时候就是对这个复合函数的链式求导。为了不把BP算法说复杂,我找了一个最简单的网络,如(图五)所示:
(图五)
(图五)中的网络只有三层:输入层X,隐藏层h和输出层y。中间有两个权重W1和W2,刚开始都是随机初始化的。神经网络的训练分为两个过程,第一过程就是从输入训练样本,层层计算到最终输出y,这个过程前向传播(forward propagation)。接着计算输出y和真实标签的差,这个可以作为简单的目标函数,我们的目标就是使得这个目标函数在所有的训练集上最小,说白了就是找个目标函数的最小值,找它就要求梯度,然后更新参数。接着就是求梯度咯,把目标函数对W2和W1求导咯,这个过程叫反向传播(back propagation)。下面(图六)简单的演示下这个两个过程:
(图六)前向传播和反向传播
注意下前向传播得到保留中间变量,比如a1,a2, 这些在反向传播中都要是用到的,整个原理也很简单,就是目标函数对各层的权重的求导,因为是复合函数,所以要链式求导。有了梯度,用 这个经典更新方式来跟新权重就行咯,其中r 是个自己设置的学习率,不要过大,大了会产生学习晃动的情况,倒三角就是梯度咯。另外输出层不一定要用(图六)中的目标函数,自己可以根据情况来指定不同的目标函数,哪怕你最后输出再加个支持向量机都行,只要你可以求导,得到梯度就行,事实上hiinton的一个弟子就在做这个事情。个人发挥自己的智慧来改进模型吧^.^,另外,卷积神经网络的参数更新过程也是类似,都免不了用BP算法来求导。下面是模仿这两个过程的代码:
- import math
- import random
- import string
- random.seed(0)
- # calculate a random number where: a <= rand < b
- def rand(a, b):
- return (b-a)*random.random() + a
- # Make a matrix (we could use NumPy to speed this up)
- def makeMatrix(I, J, fill=0.0):
- m = []
- for i in range(I):
- m.append([fill]*J)
- return m
- # our sigmoid function, tanh is a little nicer than the standard 1/(1+e^-x)
- def sigmoid(x):
- return math.tanh(x)
- # derivative of our sigmoid function, in terms of the output (i.e. y)
- def dsigmoid(y):
- return 1.0 - y**2
- class NN:
- def __init__(self, ni, nh, no):
- # number of input, hidden, and output nodes
- self.ni = ni + 1 # +1 for bias node
- self.nh = nh
- self.no = no
- # activations for nodes
- self.ai = [1.0]*self.ni
- self.ah = [1.0]*self.nh
- self.ao = [1.0]*self.no
- # create weights
- self.wi = makeMatrix(self.ni, self.nh)
- self.wo = makeMatrix(self.nh, self.no)
- # set them to random vaules
- for i in range(self.ni):
- for j in range(self.nh):
- self.wi[i][j] = rand(-0.2, 0.2)
- for j in range(self.nh):
- for k in range(self.no):
- self.wo[j][k] = rand(-2.0, 2.0)
- # last change in weights for momentum
- self.ci = makeMatrix(self.ni, self.nh)
- self.co = makeMatrix(self.nh, self.no)
- def update(self, inputs):
- if len(inputs) != self.ni-1:
- raise ValueError('wrong number of inputs')
- # input activations
- for i in range(self.ni-1):
- #self.ai[i] = sigmoid(inputs[i])
- self.ai[i] = inputs[i]
- # hidden activations
- for j in range(self.nh):
- sum = 0.0
- for i in range(self.ni):
- sum = sum + self.ai[i] * self.wi[i][j]
- self.ah[j] = sigmoid(sum)
- # output activations
- for k in range(self.no):
- sum = 0.0
- for j in range(self.nh):
- sum = sum + self.ah[j] * self.wo[j][k]
- self.ao[k] = sigmoid(sum)
- return self.ao[:]
- def backPropagate(self, targets, N, M):
- if len(targets) != self.no:
- raise ValueError('wrong number of target values')
- # calculate error terms for output
- output_deltas = [0.0] * self.no
- for k in range(self.no):
- error = targets[k]-self.ao[k]
- output_deltas[k] = dsigmoid(self.ao[k]) * error
- # calculate error terms for hidden
- hidden_deltas = [0.0] * self.nh
- for j in range(self.nh):
- error = 0.0
- for k in range(self.no):
- error = error + output_deltas[k]*self.wo[j][k]
- hidden_deltas[j] = dsigmoid(self.ah[j]) * error
- # update output weights
- for j in range(self.nh):
- for k in range(self.no):
- change = output_deltas[k]*self.ah[j]
- self.wo[j][k] = self.wo[j][k] + N*change + M*self.co[j][k]
- self.co[j][k] = change
- #print N*change, M*self.co[j][k]
- # update input weights
- for i in range(self.ni):
- for j in range(self.nh):
- change = hidden_deltas[j]*self.ai[i]
- self.wi[i][j] = self.wi[i][j] + N*change + M*self.ci[i][j]
- self.ci[i][j] = change
- # calculate error
- error = 0.0
- for k in range(len(targets)):
- error = error + 0.5*(targets[k]-self.ao[k])**2
- return error
- def test(self, patterns):
- for p in patterns:
- print(p[0], '->', self.update(p[0]))
- def weights(self):
- print('Input weights:')
- for i in range(self.ni):
- print(self.wi[i])
- print()
- print('Output weights:')
- for j in range(self.nh):
- print(self.wo[j])
- def train(self, patterns, iterations=1000, N=0.5, M=0.1):
- # N: learning rate
- # M: momentum factor
- for i in range(iterations):
- error = 0.0
- for p in patterns:
- inputs = p[0]
- targets = p[1]
- self.update(inputs)
- error = error + self.backPropagate(targets, N, M)
- if i % 100 == 0:
- print('error %-.5f' % error)
- def demo():
- # Teach network XOR function
- pat = [
- [[0,0], [0]],
- [[0,1], [1]],
- [[1,0], [1]],
- [[1,1], [0]]
- ]
- # create a network with two input, two hidden, and one output nodes
- n = NN(2, 2, 1)
- # train it with some patterns
- n.train(pat)
- # test it
- n.test(pat)
- if __name__ == '__main__':
- demo()
import math
import random
import string
random.seed(0)
# calculate a random number where: a <= rand < b
def rand(a, b):
return (b-a)*random.random() + a
# Make a matrix (we could use NumPy to speed this up)
def makeMatrix(I, J, fill=0.0):
m = []
for i in range(I):
m.append([fill]*J)
return m
# our sigmoid function, tanh is a little nicer than the standard 1/(1+e^-x)
def sigmoid(x):
return math.tanh(x)
# derivative of our sigmoid function, in terms of the output (i.e. y)
def dsigmoid(y):
return 1.0 - y**2
class NN:
def __init__(self, ni, nh, no):
# number of input, hidden, and output nodes
self.ni = ni + 1 # +1 for bias node
self.nh = nh
self.no = no
# activations for nodes
self.ai = [1.0]*self.ni
self.ah = [1.0]*self.nh
self.ao = [1.0]*self.no
# create weights
self.wi = makeMatrix(self.ni, self.nh)
self.wo = makeMatrix(self.nh, self.no)
# set them to random vaules
for i in range(self.ni):
for j in range(self.nh):
self.wi[i][j] = rand(-0.2, 0.2)
for j in range(self.nh):
for k in range(self.no):
self.wo[j][k] = rand(-2.0, 2.0)
# last change in weights for momentum
self.ci = makeMatrix(self.ni, self.nh)
self.co = makeMatrix(self.nh, self.no)
def update(self, inputs):
if len(inputs) != self.ni-1:
raise ValueError('wrong number of inputs')
# input activations
for i in range(self.ni-1):
#self.ai[i] = sigmoid(inputs[i])
self.ai[i] = inputs[i]
# hidden activations
for j in range(self.nh):
sum = 0.0
for i in range(self.ni):
sum = sum + self.ai[i] * self.wi[i][j]
self.ah[j] = sigmoid(sum)
# output activations
for k in range(self.no):
sum = 0.0
for j in range(self.nh):
sum = sum + self.ah[j] * self.wo[j][k]
self.ao[k] = sigmoid(sum)
return self.ao[:]
def backPropagate(self, targets, N, M):
if len(targets) != self.no:
raise ValueError('wrong number of target values')
# calculate error terms for output
output_deltas = [0.0] * self.no
for k in range(self.no):
error = targets[k]-self.ao[k]
output_deltas[k] = dsigmoid(self.ao[k]) * error
# calculate error terms for hidden
hidden_deltas = [0.0] * self.nh
for j in range(self.nh):
error = 0.0
for k in range(self.no):
error = error + output_deltas[k]*self.wo[j][k]
hidden_deltas[j] = dsigmoid(self.ah[j]) * error
# update output weights
for j in range(self.nh):
for k in range(self.no):
change = output_deltas[k]*self.ah[j]
self.wo[j][k] = self.wo[j][k] + N*change + M*self.co[j][k]
self.co[j][k] = change
#print N*change, M*self.co[j][k]
# update input weights
for i in range(self.ni):
for j in range(self.nh):
change = hidden_deltas[j]*self.ai[i]
self.wi[i][j] = self.wi[i][j] + N*change + M*self.ci[i][j]
self.ci[i][j] = change
# calculate error
error = 0.0
for k in range(len(targets)):
error = error + 0.5*(targets[k]-self.ao[k])**2
return error
def test(self, patterns):
for p in patterns:
print(p[0], '->', self.update(p[0]))
def weights(self):
print('Input weights:')
for i in range(self.ni):
print(self.wi[i])
print()
print('Output weights:')
for j in range(self.nh):
print(self.wo[j])
def train(self, patterns, iterations=1000, N=0.5, M=0.1):
# N: learning rate
# M: momentum factor
for i in range(iterations):
error = 0.0
for p in patterns:
inputs = p[0]
targets = p[1]
self.update(inputs)
error = error + self.backPropagate(targets, N, M)
if i % 100 == 0:
print('error %-.5f' % error)
def demo():
# Teach network XOR function
pat = [
[[0,0], [0]],
[[0,1], [1]],
[[1,0], [1]],
[[1,1], [0]]
]
# create a network with two input, two hidden, and one output nodes
n = NN(2, 2, 1)
# train it with some patterns
n.train(pat)
# test it
n.test(pat)
if __name__ == '__main__':
demo()
转载请注明来源: http://blog.csdn.net/marvin521/article/details/9886643
参考文献:
[1] Learning From Data. Yaser S.Abu-Mostafa
[2] machine learning.Andrew Ng