chunked/nltk_chunking/model_doc_vit_mae/chunk_6.txt

Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task.