New paper: "Improving Vision Transformer Efficiency and Accuracy by Learning to Tokenize"

[JackKelly]

Detailed Description

Very interesting new paper from Google:

• Blog: Improving Vision Transformer Efficiency and Accuracy by Learning to Tokenize
• Paper: TokenLearner: What Can 8 Learned Tokens Do for Images and Videos?

Basically: A way to dynamically pick which "patches" of images / videos to attend to.

This could be a really nice way of including much larger input images (which is important for longer-time-horizon nowcasting).