The pre-training
objective is to match the prototypes assigned to the unmasked views of the images to that of the masked views of the same images.