Training a model on unlabeled data using the data itself to create learning signals, without manual annotations.