Training independent subnetworks for robust prediction