Cascade-Correlation learning method has conventionally been used for training feedforward multilayer networks with sigmoidal activation in the hidden units. We have developed a modification of the Cascade-Correlation method and applied it successfully in training radial basis function networks. This method works in a similar way as the orthogonal least squares method which has been utilized for training radial basis function networks. Our experiments show that with the modified Cascade-Correlation training approximately the same error level can be reached but significantly faster than with the orthogonal least squares training.
- neural networks
- radial basis function network