Fig. 2From: Comparison of semi-supervised deep learning algorithms for audio classificationDCT workflow. Each model is trained on its own labeled samples xi, unlabeled samples xu, and the adversarial examples generated by the other model. Model f makes predictions on x1 and \(x^{g}_{4}\), and model g on x2 and \(x^{f}_{3}\). In our DCT+mixup variant, mixup is used on the unlabeled samples onlyBack to article page