Generalization in cascade-correlation networks

Two network construction algorithms are analyzed and compared theoretically as well as empirically. The first algorithm is the cascade correlation learning architecture proposed by S. E. Fahlman (1990), while the other algorithm is a small but striking modification of the former. Fahlman's algorithm builds multilayer feedforward networks with as many layers as the number of added hidden units, while the other algorithm operates with just one layer of hidden units. This implies that their computational capabilities and the representation of the generalizations they deal with are quite diverse, and it is demonstrated how the generalization ability of the networks generated by Fahlman's algorithm is outperformed by the networks built by the new algorithm.<>