理解神经网络，从简单的例子开始（2）使用python建立多层神经网络

这篇文章将讲解如何使用python建立多层神经网络。在阅读这篇文章之前，建议先阅读上一篇文章:理解神经网络，从简单的例子开始。讲解的是单层的神经网络。如果你已经阅读了上一篇文章，你会发现这篇文章的代码和上一篇基本相同，理解起来也相对容易。

上一篇文章使用了9行代码编写一个单层的神经网络。而现在，问题变得更加复杂了。下面是训练输入数据和训练输出数据，如果输入数据是[1，1，0]，最后的结果是什么呢？

从上面的输入输出数据可以找出规律：第一和第二列的值“异或”之后得到output的值，而第三列没有关系。所谓的异或，就是相同为0，相异为1.
所以，当输入数据为[1，1，0]时，结果为0。
但是，这个在单层网络节点中是很难实现的，因为input和output之间没有一对一的对应关系。所以，可以考虑使用多层的神经网络，或者说增加一个隐藏层，叫layer1，它能够处理input的合并问题。

上图就是新的神经网络图。蓝线表示两个神经元之间连接的神经突触。这个图片可以使用这个开源代码自动生成： https://github.com/miloharper/visualise-neural-network。

从图中可以看出，第1层的输出进入第2层。现在神经网络能够处理第一层的输出数据和第二层（也就是最终输出的数据集）之间的关系。随着神经网络学习，它将通过调整两层的权重来放大这些相关性。

实际上，图像识别和这个很相似。比如下面图片中，一个像素和苹果没有直接的关系，但是许多像素组合在一起就能够构成一个苹果的样子，也就产生了关系。

这里通过增加更多层神经来处理“组合”问题，便是所谓的深度学习了。下面是多层神经网络的python代码。解释会在代码注释中和代码后面。

from numpy import exp, array, random, dot


class NeuronLayer():
    def __init__(self, number_of_neurons, number_of_inputs_per_neuron):
        self.synaptic_weights = 2 * random.random((number_of_inputs_per_neuron, number_of_neurons)) - 1


class NeuralNetwork():
    def __init__(self, layer1, layer2):
        self.layer1 = layer1
        self.layer2 = layer2

    # The Sigmoid function, which describes an S shaped curve.
    # We pass the weighted sum of the inputs through this function to
    # normalise them between 0 and 1.
    def __sigmoid(self, x):
        return 1 / (1 + exp(-x))

    # The derivative of the Sigmoid function.
    # This is the gradient of the Sigmoid curve.
    # It indicates how confident we are about the existing weight.
    def __sigmoid_derivative(self, x):
        return x * (1 - x)

    # We train the neural network through a process of trial and error.
    # Adjusting the synaptic weights each time.
    def train(self, training_set_inputs, training_set_outputs, number_of_training_iterations):
        for iteration in range(number_of_training_iterations):
            # Pass the training set through our neural network
            output_from_layer_1, output_from_layer_2 = self.think(training_set_inputs)

            # Calculate the error for layer 2 (The difference between the desired output
            # and the predicted output).
            layer2_error = training_set_outputs - output_from_layer_2
            layer2_delta = layer2_error * self.__sigmoid_derivative(output_from_layer_2)

            # Calculate the error for layer 1 (By looking at the weights in layer 1,
            # we can determine by how much layer 1 contributed to the error in layer 2).
            layer1_error = layer2_delta.dot(self.layer2.synaptic_weights.T)
            layer1_delta = layer1_error * self.__sigmoid_derivative(output_from_layer_1)

            # Calculate how much to adjust the weights by
            layer1_adjustment = training_set_inputs.T.dot(layer1_delta)
            layer2_adjustment = output_from_layer_1.T.dot(layer2_delta)

            # Adjust the weights.
            self.layer1.synaptic_weights += layer1_adjustment
            self.layer2.synaptic_weights += layer2_adjustment

    # The neural network thinks.
    def think(self, inputs):
        output_from_layer1 = self.__sigmoid(dot(inputs, self.layer1.synaptic_weights))
        output_from_layer2 = self.__sigmoid(dot(output_from_layer1, self.layer2.synaptic_weights))
        return output_from_layer1, output_from_layer2

    # The neural network prints its weights
    def print_weights(self):
        print( "    Layer 1 (4 neurons, each with 3 inputs): ")
        print( self.layer1.synaptic_weights)
        print( "    Layer 2 (1 neuron, with 4 inputs):")
        print( self.layer2.synaptic_weights)

if __name__ == "__main__":

    #Seed the random number generator
    random.seed(1)

    # Create layer 1 (4 neurons, each with 3 inputs)
    layer1 = NeuronLayer(4, 3)

    # Create layer 2 (a single neuron with 4 inputs)
    layer2 = NeuronLayer(1, 4)

    # Combine the layers to create a neural network
    neural_network = NeuralNetwork(layer1, layer2)

    print ("Stage 1) Random starting synaptic weights: ")
    neural_network.print_weights()

    # The training set. We have 7 examples, each consisting of 3 input values
    # and 1 output value.
    training_set_inputs = array([[0, 0, 1], [0, 1, 1], [1, 0, 1], [0, 1, 0], [1, 0, 0], [1, 1, 1], [0, 0, 0]])
    training_set_outputs = array([[0, 1, 1, 1, 1, 0, 0]]).T

    # Train the neural network using the training set.
    # Do it 60,000 times and make small adjustments each time.
    neural_network.train(training_set_inputs, training_set_outputs, 60000)

    print ("Stage 2) New synaptic weights after training: ")
    neural_network.print_weights()

    # Test the neural network with a new situation.
    print ("Stage 3) Considering a new situation [1, 1, 0] -> ?: ")
    hidden_state, output = neural_network.think(array([1, 1, 0]))
    print (output)

上图是深度学习的计算周期

和上篇博客中的单层神经网络相比，这里是多层，也就是增加了隐藏层。当神经网络计算第二层的error误差值的时候，会把这个error向后面一层也即是第一次传播，从而计算并调整第一层节点的权值。也就是，第一层的error误差值是从上一层也即是第二层所传播回来的值中计算得到的，这样就能够知道第一层对第二层的误差有多大的贡献。

运行上面的代码，会得到下面所示的结果。

Stage 1) Random starting synaptic weights: 
    Layer 1 (4 neurons, each with 3 inputs): 
[[-0.16595599  0.44064899 -0.99977125 -0.39533485]
 [-0.70648822 -0.81532281 -0.62747958 -0.30887855]
 [-0.20646505  0.07763347 -0.16161097  0.370439  ]]
    Layer 2 (1 neuron, with 4 inputs):
[[-0.5910955 ]
 [ 0.75623487]
 [-0.94522481]
 [ 0.34093502]]
Stage 2) New synaptic weights after training: 
    Layer 1 (4 neurons, each with 3 inputs): 
[[ 0.3122465   4.57704063 -6.15329916 -8.75834924]
 [ 0.19676933 -8.74975548 -6.1638187   4.40720501]
 [-0.03327074 -0.58272995  0.08319184 -0.39787635]]
    Layer 2 (1 neuron, with 4 inputs):
[[ -8.18850925]
 [ 10.13210706]
 [-21.33532796]
 [  9.90935111]]
Stage 3) Considering a new situation [1, 1, 0] -> ?: 
[ 0.0078876]

这篇文章是本人根据这篇博客写的。一定程度上算是在做翻译的工作。本人特别推荐该源码的风格，一看就知道这是编程经验丰富之人才能写出来的代码，清晰明了，看起来特别舒服。

结束！ 转载请标明出处，感谢！