We make neural network training cheaper and more accurate by progressively dropping parts of the data after each epoch.