CLIP is able to predict this by learning the cosine similarity between image and text feature for batches of training data. This is shown in the contrastive pre-training portion of Figure 1 where the ...