Since our data is unlabeled, won't there be pairs of images in which both are different cats? Hence wouldn't assigning them as "negative pairs" not teach the model that they are both cats?