Abstract: Deep Belief Network（DBN） is a deep architecture that consists of several Restricted Boltzmann Machines（RBM）. Generally the inputs of RBM are binary vector which leads to the information loss and in turn degrades the performance of classification. For the problem above, a DBN classification algorithm for numerical attribute is proposed through scaling the input into the interval between 0 and 1 with adding noise to sigmoid units, and achieving classification with one Gaussian hidden node on the top-level RBM. DBN can be used as feature extraction method as well as neural network with initially learned weights. DBN should have a better performance than the traditional neural network due to the initialization of the connecting weights rather than just using random weights in neural network. Experiments conducted on the dataset from UCI show that the proposed algorithm has a better accuracy than the traditional classification algorithm like SVM.

Key words: numerical attribute, classification, Deep Belief Network（DBN）, associate memory

摘要： 深度置信网络是个包含多个受限玻尔兹曼机的深层架构。针对深度置信网络分类时由于受限玻尔兹曼机的输入一般是二值向量而造成的信息的丢失从而使分类效果降低的问题，提出了通过在sigmoid单元中增加噪声来将输入缩放到[0，1]区间，使用带有一个高斯隐藏节点的顶层受限玻尔兹曼机实现分类功能的一种数值属性深度置信网络分类方法。深度置信网络和受限玻尔兹曼机可以作为特征提取方法也可以认为是带有训练的初始权值的神经网络。由于连接权值的初始化而不仅仅是神经网络的随机权值，深度置信网络分类应该比原有的传统的神经网络分类拥有更好的性能。在UCI的多个数据集上进行对比验证，实验结果表明深度置信网络分类方法比传统的SVM算法拥有更高的准确性。

关键词: 数值属性, 分类, 深度置信网络, 联想记忆