Semi-supervised learning uses labeled and unlabeled data from known classes for training, assuming the test data contains only those classes. However, in real-world scenarios, new classes can appear.