Masked Autoencoder

Improved representations since images have heavy spatial redundancy (

Image Locality). Therefore, learning to recover a missing patch from neighboring patches is helpful for high-level understanding.