Abstract: In recent years,the deep learning to get a successful application in image,voice,video and other unstructured data,has become a hot topic of machine learning and data mining.As a supervised learning model,the successful deep learning applications often require a larger set of high-quality training.Based on this situation,we studied deep belief network composed of more restricted Boltzmann machines,and combined with the thought of semi-supervised learning,we used smaller training set to improve the classification accuracy of depth network model.We used Knn,SVM and pHash three methods to study the non-labeled data set.And the result shows that the semi-supervised deep belief networks increases image classification accuracy by about 3% compared with the traditional network with more restricted Boltzmann machine.

Key words: Semi-supervised,Deep belief networks,Restricted boltzmann machine