The pre-training objective is relatively simple:
by masking a large portion (75%) of the image patches, the model must reconstruct raw pixel values.