Robust Multiple-Instance Learning with Superbags