The study states that the distribution shift between training data and unlabeled data was reduced using a shift-robust regulariser.