An active pattern set strategy for enhancing generalization while improving backpropagation training efficiency

An active pattern set strategy is presented which provides a simple approach to reducing the training time for large pattern sets which contain redundant information. This approach to neural network training addresses the problem of scalability. The strategy involves the systematic removal of patterns from the training set as they are learned, thus allowing the computational resources to be concentrated on those patterns which are difficult to learn. A comparative study with standard backpropagation training indicates lower training times for the active pattern set strategy with improvements of up to a factor of three. The improvement in convergent rate did not result in a degradation of the ability of the trained network to generalize to patterns not included in the training set. Total training time in an EKG (electrocardiography) rhythm application has been reduced.< >