Consider a classifier trained on the 60k MNIST images (no augmentation). We test it on images where the pixels have been shifted. We expect the test accuracy to be poor. Is this an example of distribution shift? Please provide a reason.