A training approach where a model learns patterns from unlabeled data by creating its own learning targets, such as predicting hidden parts of the input.
World knowledge accuracy, recall of facts and relationships