深度信念网络(DBN)网络层次数量的研究及应用

Research and Application of the Level of the Deep Belief Network( DBN)

针对DBN网络隐含层层数难以选择的问题,从通信原理、信息理论以及实验数据等多方面出发,研究了深度信念网络(DBN)隐含层的层次趋势问题。根据各个隐含层输出层不同类图片的互相关系数之间的关系,提出了一种根据互相关系数确定网络深度的方法,证明了当深度学习时,隐含层输出的样本之间的互相关系数达到1(0)或^(-1)时,或者样本之间的互相关系数不再改变时,进一步增加层次对提高分类正确率是没有帮助的。在训练的过程中随机的选取图片,使其更具有普适性。手写体数据库实验和应用于CIFAR^(-1)0图像库的数据表明,该方法能够有效提高训练速度。

Aiming at the problem that the hidden layer number of DBN network is hard to choose,from the aspects of communication theory,information theory and experimental data,the problem of the level trend of the hidden layer of deep belief network( DBN) is studied. According to the mutual correlation coefficient among different class of pictures of each hidden layer and output layer,a method is put forward to determine the depth of the network according to the correlation coefficient. When the cross-correlation coefficient of the hidden layers of the samples reaches 1( 0) or^(-1) through deep learning or the cross-correlation coefficient between samples is no longer change. It Can't help to improve the correct classification rate through increase the level. Selected pictures randomly in the training process,making it more universality. Experiments of handwritten database and the data from the CIFAR^(-1)0 image database show that the method can effectively improve the training speed.