Contrastive learning
It’s like asking, “can you distinguish the real thing from the fakes?” Being able to answer this question requires a lot of understanding about the data.
It allows us to simply use the existing data (rather than building a training set purposely) to train powerful models.