Unsupervised Alternating Projection Neural Network with Convex Constraint
Hengqing Tong, Tianzhen Liu, Yang Liu, Qiaoling Tong · 2007
Alternating projection neural networks(APNN) have been researched for many years. This paper proposes a kind of APNN which is also a unsupervised neural net- work(UAPNN) with convex constraint. A linear regression model with unknown dependent variable and constrained regression coefficients is constructed. The dependent variables of the model is unknown, but it can be expressed as a linear combination according to the evaluation groups. The main characteristics of the samples are learned after training. In order to realize unsupervised learning of the neural network with convex constraint, an iterative computation method that makes use of alternating projection between two convex sets is proposed. The final example shows that the computation converges very fast.Our work may enrich the theory of neural network and also expand the evaluation method.