Masked Autoencoders Are Scalable Vision Learners

lastmod 2024-02-06created 2024-02-062 edits

About

• (Paper study) (Detailed review) Masked Autoencoders are Scalable Vision Learners :: Blue collar Developer

Masked Autoencoders Are Scalable Vision Learners (link) is a paper from Facebook AI Research (still listed as Facebook for now, but it will probably change to Meta?), and it appears to be currently submitted to CVPR 2022 and under review.

In short, the Masked autoencoder is about the idea that

Masked patch를 복원하는 task로 학습한 것을 pretrained weight로 삼아서 Downstream task로 Finetuning했을 때 예측 성능이 더 좋다.

That is the gist.

Just as BERT appeared after the Transformer in language models, I think MAE is the BERT equivalent for the Vision Transformer (ViT).

Documentation

[2111.06377] Masked Autoencoders Are Scalable Vision Learners

https://arxiv.org/abs/2111.06377

See also

• Autoencoder
• facebook