A dynamic node architecture scheme for backpropagation neural networks

Typically, artificial neural network (ANN) training schemes require network architectures to be set before training. However, the Teaming speed and generalization characteristics of ANNs are dependent on their architectures. Thus, the viability of a specific architecture can only be evaluated after training. This work seeks to reduce the dependence of ANN capabilities on the preselection of network architectures. The present work describes an ANN dynamic node architecture (DNA) scheme which determines the appropriate number of nodes for a given network by defining an importance function which assigns an importance to each node in the network. Optimizing the network architecture becomes part of the training objective. The backpropagation learning algorithm has been implemented with this new DNA scheme.