General-to-Specialized Analysis Based on Deep Belief Network

Recently, the deep neural network has achieved great performance in many areas. During analysis, all learned features are used at once, some of which could bring negative affect to specific classes. Recently, cognitive studies show that a human visual cognition process is hierarchical and dynamic, i.e., when meeting different targets, human brain intends to pay attention to different parts. Therefore, in this paper, we introduce this kind of mechanism into deep belief network (DBN) and propose a new general-to-specialized algorithm. Firstly, hierarchical knowledge networks are constructed based on the original learned DBN through pruning and retraining. Because these networks are learned for different discriminate ability, we call them as the general network and the specialized network separately. Secondly, a general-to-specialized analysis method is proposed which is proved theoretically. When predicting the class of an input sample, we select the corresponding specialized networks according to the preliminary analysis result and then make in-depth analysis. Experiments on four benchmark datasets are performed to test the proposed algorithm. The results show that our algorithm is feasible, valid and robust.