A comparative study of output representation schemes for multilayer neural networks

In this paper, we compare the 1-out-of-N representation scheme with three distributed ones, namely binary, Gray, and simple-sum. We put the emphasis on the training time, learning accuracy, and generalization capability. In order to evaluate the performance of these schemes, three multilayer neural networks (multilayer perceptron, multilayer quadratic perceptron, and multi-sieving network) are used to learn the vowel recognition and image segmentation problems.