In March, there was a lot of news about the use of self-learning for computer vision tasks. The main problem that self-learning models try to solve is to perform tasks without relying on carefully selected and labeled datasets. FAIR and Microsoft presented several studies on this topic at once.
SEER
Accessibility: blog post / article
Facebook SEER (SElf-supERvised), 1.3 . Instagram 84.2% Imagenet.
— SwAV - , RegNet VISSL — - PyTorch. 60 self-supervised , .
Learning from videos
Facebook , , . Instagram Reels, , . , 20%.
Time-Space Transformer
TimeSformer . self-attention NLP. 3D , . - . self-attention , . , , , , .
Swin Transformer
Microsoft , . , , . , , . , PyTorch.
LightningDOT
. - .
Microsoft , . . SOTA- Flickr30K COCO.
Involution
self-attention . Involution — , .
TagMe
, . . GPS- , . , . , .
PF-AFN
, GAN- . . , , . , , .
Latent Composition
GAN , . MIT , . GAN , . .
Infinite Images
MIT , «» . , . , , . , Infinite nature, , .
GANsformer
StyleGAN , . , . , , .
FaceBlit
GAN . , . , , - , .
CLIP
, CLIP Open AI , iPod .
Open AI « » .
, CLIP , , , — , . , Open AI — , , .
, , , «» , , . , CLIP . . , , . , , , .
P.S.
, - , .
, :
source code and collaboration of the Infinite Nature project from the January selection ;
an interactive demo for Monster Mash from the September issue ;
GPT-Neo , an unofficial implementation of the 2.7 billion parameters GPT-3 from Eleuther.ai.
The open source community continues to delight, and in March even a service appeared through which you can file a complaint about non-reproducible research results.
That's all, thank you for your attention and see you next month!