5. MAE: Stop Matching Views and Reconstruct Missing Content
Masked Autoencoders take a different route.
Instead of creating two augmented views and forcing their representations to agree, MAE masks a large fraction of the image.
Typically around (75%) of the patches are removed.
Only visible patches enter the encoder. A lightweight decoder receives the encoded visible patches plus mask tokens for missing positions and reconstructs the missing pixels.
The objective is roughly:
This changes the collapse problem. A constant representation cannot reconstruct the different pixel targets of many different images.
MAE therefore changes the learning problem itself:
Instead of defining invariance between different views, predict the content that has been removed.
But this introduces another question:
Do we really need to reconstruct every pixel?