A semi-supervised learning technique where a teacher model is used to generate pseudo-labels on unlabeled data, then a student model is trained on both labeled and unlabeled data. The student model is encouraged to have higher capacity and during its training, random noise is added to its weights.