SIGMOID LOSS FOR LANGUAGE IMAGE PRETRAINING

Summary from the TLDR AI newsletter

CLIP was somewhat revolutionary for computer vision. It has some extremely clever algorithmic tricks (e.g., large in-batch negatives) that make it perform well. This paper questions many of those tricks and finds a simpler training setup that outperforms CLIP for the same task.