| Popis: |
A method for improving back-propagation-based training by augmenting the output patterns with additional relevant information is presented. It is suggested that the augmented output provides additional constraints that more precisely specify the allowable function. This results in faster training and better generalization. Improvement depends on the size of the intersection of the two classes of possible mapping functions. In so far as the intersection is not empty, performance may improve. An empirical advantage of the extra-output technique is that after a network has been trained to realize a desired function, the extra output units may be detached. The resulting network computes more rapidly in that it has fewer connections to manipulate |
