Segregation algorithm: splitting the training set improves learning

Summary form only given. A learning algorithm for feedforward backpropagation networks is proposed that improves the net´s ability to generalize or to learn the training set compared to standard backpropagation, using an equal number of learning steps. Thus the algorithm needs fewer learning steps to obtain the same percentage of correct responses in the training set. The proposed algorithm makes an estimate of the required number of units and increases it when necessary. It takes into account the distribution of patterns among classes and adapts its topology accordingly. Execution time is significantly reduced as fewer weight updates per learning iteration are required.<>