Research of DBN PLSR algorithm Based on Sparse Constraint

Abstract

DBN is a generative model based on unsupervised learning, with strong computing and information processing capabilities. But at the same time, there are some drawbacks: the model is constructed through intensive expression, which leads to relatively low computing performance of the network. The network optimization method based on the BP algorithm is easy to fall into a local minimum, which makes DBN fine-tuning accuracy is reduced. In order to obtain a DBN that is efficient and can avoid local optimization, the paper designs a DBN based on adaptive sparse representation and partial least square regression (PLSR) fine-tuning. First, two regularization factor terms are introduced to punish the densely expressed connection characteristics, thereby constructing an sparse RBM. Secondly, PLSR method is adopted instead of the BP algorithm, and a PLSR model is established between every two layers from the output layer to the input layer. The experiment proved the effectiveness of optimized DBN in improving network performance and learning performance. Project Supported by Natural Science Basic Research Program of Shaanxi (Program No.2020JQ-788). Project Supported by Natural Science Basic Research Program of Shaanxi (ProgramNo.2020JM-542).