Training ML models on unlabeled data by automatically generating labels or tasks from the data itself.