Learning Transferable Visual Models From Natural Language Supervision

CLIP論文, OPENAIのRadford

Introduction
Mori et al improved content-based image retrieval.
Joulin modernized the retrieval.

Approach
100 million photos to 6 to only 15 million photos.
Publicly available sources on the Internet.
WebImageText

N labels are used for CLIP and the famous contrastive learning pre-training method is also utilized.

Data overlap analysis
Pre-training on a very large internet dataset
It can be an overlap between training images and test images.

Conclusion
This paper explored whether transferring in NLP is possible to another domain.