A machine learning approach where models learn from unlabeled data by creating their own supervision signals, such as predicting missing parts of input. Bridges supervised and unsupervised learning methods.