Design artificial neural networks based on the principle of divide-and-conquer

A split-and-merge growth learning algorithm that builds a feedforward network of threshold logic units with two hidden layers to classify any given (nonlinearly) separable training patterns is proposed. The algorithm simultaneously learns the structure and connection weights of the network. Fast convergence in finite steps to the correct classification is guaranteed. It is based on the principle of divide-and-conquer. The ways in which this algorithm differs from other growth algorithms is discussed