Masked Autoencoder (MAE)

What is Masked Autoencoder (MAE)?

A vision model trained by masking out large chunks of an image (e.g., 75%) and forcing the neural network to reconstruct the missing pixels.

Where did the term "Masked Autoencoder (MAE)" come from?

BERT-like pre-training for images.

How is "Masked Autoencoder (MAE)" used today?

State-of-the-art for visual representation learning.

Related Terms

• Self-Supervised Learning
• Vision Transformer (ViT)