Researchers report artificial neural networks contain subnetworks that are considerably smaller, yet capable of being trained to make equally accurate predictions. Many subnetworks can learn to make accurate predictions much faster than the original networks.