RankAlign

RankAlign method aligns vision-language models using rank transformation

Researchers have introduced RankAlign, a novel method for aligning vision and language representations without requiring cross-modal supervision. This approach transforms within-modality similarity kernels into ordinal …