ALADIN: algorithms for learning and architecture determination

The development of fully autonomous algorithms which are capable of selecting and training the feedforward neural network with the best performance for a given application is presented. The proposed learning algorithms determine the architecture of multilayered neural networks while performing their training. The architecture of the networks is determined during the training by inactivating the redundant hidden units on the basis of a criterion relating to the effect of each hidden unit on the performance of the network. In addition to the algorithms based on the least squares criterion frequently used for the training of neural networks, fast algorithms based on a novel generalized criterion which accelerates the training of neural networks are developed. Several experiments verify that the proposed algorithms provide the simplest neural networks with the highest generalization efficiency