Fig. 4From: Comparison of semi-supervised deep learning algorithms for audio classificationRMM workflow. One weak and K strong augmentations are applied to the unlabeled data xu. The weakly augmented unlabeled data \(x^{\prime }_{u}\) are used to create pseudo labels \(\hat{y}_{u}\). The first batch of strongly augmented unlabeled data \(x^{\prime }_{u_{i}}, i=1\) is used in the unsupervised loss component \(L_{u_{1}}\) (using the pseudo-labels \(\hat{y}_{u}\))Back to article page