Dynamical configuration of neural network architectures

A dynamical configurable architecture for feedforward artificial neural networks (ANNs) is proposed. A dynamical configuration rule based on a general topological structure for feedforward neural networks and an adaptive learning algorithm are presented. The two combined provide an automated paradigm for synthesis of feedforward ANNs that has the potential to generate the optimal ANN representations for arbitrary training samples. Since the size of the architecture is determined by the dynamical configuration rule autonomously, this paradigm is advantageous in terms of convenience of architectural realization and reduction of computational time