“A key challenge in training models on unlabeled visual data is linking abstract concepts to what is seen in an image without explicit language annotations.”