Heuristic configuration of hidden units, in backpropagation neural networks

Summary form only given, as follows. For optimum statistical classification and generalization with single hidden-layer backpropagation neural network models, two tasks must be performed: (a) learning the best set of weights for a network of k hidden units and (b) determining k, the best complexity fit. Two approaches to learning have been contrasted: (a) standard backpropagation as applied to a series of networks with different numbers of hidden units; and (b) a heuristic cascade-correlation approach that quickly and dynamically learns and configures a network. Four real-world statistical applications were considered. On these examples, the backpropagation approach yielded somewhat better results, but with far greater computation times. The best k´s for the two approaches were quite similar, suggesting a hybrid approach that chooses k by cascade-correlation, and optimizes the weights by backpropagation