Rhea Chowers

Researchers find modality gap in AI models can improve robustness

Researchers have investigated the modality gap in multi-modal models like CLIP, observing that images and texts often occupy separate distributions in the shared embedding space. This paper demonstrates that this gap ca…