A SIMPLE APPROACH TO VISION-LANGUAGE CLASSIFICATION

Summary from the TLDR AI newsletter

SuperClass is a straightforward classification model for vision-language tasks that skips the need for a text encoder, unlike contrastive models like CLIP. SuperClass avoids complex text filtering and large batch requirements by directly using tokenized raw text as classification labels.