人工神经网络(Artificial Neural Network, ANN)[1]具有很广的应用。理论上来说,采用多层的神经网络能够逼近任何的连续函数,而不管该函数是否平滑。自从SVM出现后,大家都快忘了还有ANN这种东东。但是近几年随着deep learning技术的兴起,大牛们又重新开始关注神经网络了。在这里,演示一种个人认为最简单的三层神经网络。它包含一个输入层(即数据本身),一个中间层(又称隐含层,hidden layer), 和一个输出层。其中,输入层的大小即为数据的维度,中间层大小可调,输出层只包括一个神经元。当该神经元输出接近1时,判断为第一类,输出为0时判断为第二类。所有激活函数均采用sigmoid函数:
神经网络的训练最常用的方法是误差反传播(back-propagation, BP),本质上其实是梯度下降算法。其推导过程需要用到高数中的链式法则(chain rule),在此不作推导,只在代码中给出结果。有不明白的小盆友可以发email给我。训练部分的代码如下:
%function output = ANN_Training(X, t, H, max_iter)
%Train ANN with one hidden layer and one output unit for classification
%input :
%X: attributes. Every column represents a sample.
%y: target. should be 0 or 1. length(y) == size(X,2) is assumed.
%H: size of hidden layer.
%max_iter: maximum iterates
%tol: convergence tolerate
%output:
% output: a structure containing all network parameters.
%Created by Ranch Y.Q. Lai on Mar 22, 2011
%[email protected]
%Train ANN with one hidden layer and one output unit for classification
%input :
%X: attributes. Every column represents a sample.
%y: target. should be 0 or 1. length(y) == size(X,2) is assumed.
%H: size of hidden layer.
%max_iter: maximum iterates
%tol: convergence tolerate
%output:
% output: a structure containing all network parameters.
%Created by Ranch Y.Q. Lai on Mar 22, 2011
%[email protected]