A new node selection method of Deep Belief Network based on similarity

摘要

Research on the node selection problem in Deep Belief Network (DBN). A new node selection model based on similarity is proposed to select nodes in the pre-training process of DBN, which is an unsupervised process. Firstly, extracting the nodes feature from pre-training weight matrix, and then comparing them each other to get the similarity matrix, and using the similarity matrix to select nodes and reusing the nodes. Finally, the test on MINIST data set proves that the new model has better performance.