Dynamic neural networks with the use of divide and conquer

An algorithm called divide and conquer neural networks that creates a feedforward neural network during training based upon the training examples is described. In addition to learning the weights for connections, it learns an architecture that enables it to learn the examples. Training is done on the inputs to one cell at a time with learned weights being frozen. Error is never propagated backwards through a hidden cell. Examples of the algorithm´s performance on the exclusive-OR and the Iris plant data, which contain two nonlinearly separable classes, are given. The results show that this algorithm can effectively learn a viable architecture in which training examples may be encoded and generalized for later use in classification