Semi-supervised learning, i.e. jointly learning from labeled and unlabeled samples, is an active research topic due to its key role on relaxing human supervision. In the context of image classification, recent advances to learn from unlabeled samples are mainly focused on consistency regularization methods that encourage invariant predictions for different perturbations of unlabeled samples. We, conversely, propose to learn from unlabeled data by generating soft pseudo-labels using the network predictions. We show that a naive pseudo-labeling overfits to incorrect pseudo-labels due to the so-called confirmation bias and demonstrate that mixup augmentation and setting a minimum number of labeled samples per mini-batch are effective regularization techniques for reducing it. The proposed approach achieves state-of-the-art results in CIFAR-10/100, SVHN, and Mini-ImageNet despite being much simpler than other methods. These results demonstrate that pseudo-labeling alone can outperform consistency regularization methods, while the opposite was supposed in previous work. Source code is available at https://git.io/fjQsC.

The ventral midbrain (vMb) is organized into distinct anatomical domains and contains cohorts of functionally distinct subtypes of midbrain dopamine (mDA) neurons. We tested the hypothesis that genetic history and timing of gene expression within mDA neuron progenitors impart spatial diversity. Using Genetic Inducible Fate Mapping to mark the Sonic hedgehog (Shh) and Gli1 lineages at varying embryonic stages, we performed a quantitative and qualitative comparison of the two lineages’ contribution to the mDA neuron domains. Dynamic changes in Shh and Gli1 expression in the vMb primordia delineated their spatial contribution to the E12.5 vMb: Both lineages first contributed to the medial domain, but subsequently the Gli1 lineage exclusively contributed to the lateral vMb while the Shh lineage expanded more broadly across the vMb. The contribution of both lineages to the differentiated mDA neuron domain was initially biased anteriorly and became more uniform across the anterior/posterior vMb throughout development. Our findings demonstrate that the early Shh and Gli1 lineages specify mDA neurons of the substantia nigra pars compacta while the late Shh and Gli1 lineages maintain their progenitor state longer in the posterior vMb to extend the production of mDA neurons in the ventral tegmental area. Together, our study demonstrates that the timing of gene expression along with the genetic lineage (Shh or Gli1) within the neural progenitors segregate mDA neurons into distinct spatial domains.

No abstract