Improved neural network training using redundant structure

It is a common understanding in neural network research and applications that a network with fewer redundant nodes is more reliable. This paper argues that a redundant network structure approach improves the learning process of neural networks. This redundant structure is shown to be free from extra parameters and hence does not introduce additional uncertainty. Using a small partition problem, the training results of standard BP networks are compared with those networks with a redundant structure. The comparison shows that a redundant structure does not necessarily always have a negative effect, and as a result it is possible to help a neural network obtain better performance.