Facebook AI Introduces Masked Autoencoders as Scalable Vision Learners
Facebook AI Research demonstrates that masked autoencoders bring the success of NLP self-supervised learning to computer vision. The new approach efficiently trains large models by reconstructing heavily masked images.
A Facebook AI Research (FAIR) team introduces masked autoencoders (MAE) as highly scalable self-supervised learners for computer vision, bringing the proven success of NLP methods like BERT to the visual domain. While masked language modeling achieves astonishing results in natural language processing, the progress of autoencoding methods in computer vision significantly lags behind until now.
The researchers address the differences between vision and language domains, noting distinct architectures, varying information densities, and the different roles of autoencoder decoders. To overcome these challenges, the team develops a simple yet effective MAE method that randomly masks patches from an input image and tasks the model with reconstructing the missing pixels.
This approach utilizes an asymmetric encoder-decoder architecture where the encoder processes only visible patches while a lightweight decoder handles the reconstruction. By masking a remarkably high proportion of the input image—up to 75 percent—the framework creates a meaningful self-supervisory task that enables efficient and effective training of large computer vision models.