They exploit the momentum of the gradients to figure out how speedy weights modify throughout instruction and prune these that do not alter considerably.