Deep learning has revolutionized image analysis and natural language processing with remarkable accuracies in prediction tasks, such as image labeling or word identification. The origin of this revolution was arguably the deep learning approach by the Hinton lab in 2012, which halved the error rate of existing classifiers in the then 2year-old ImageNet database 1 . In hindsight, the combination of algorithmic and hardware advances with the appearance of large and well-labeled datasets has led up to this seminal contribution. The emergence of large amounts of data from single-cell RNA-seq and the recent global effort to chart all cell types in the Human Cell Atlas has attracted an interest in deep-learning applications. However, all current approaches are unsupervised, i.e., learning of latent spaces without using any cell labels, even though supervised learning approaches are often more powerful in feature learning and the most popular approach in the current AI revolution by far. Here, we ask why this is the case. In particular we ask whether supervised deep learning can be used for cell annotation, i.e. to predict cell-type labels from single-cell gene expression profiles. After evaluating 6 classification methods across 14 datasets, we notably find that deep learning does not outperform classical machine-learning methods in the task. Thus, cell-type prediction based on gene-signature derived celltype labels is potentially too simplistic a task for complex non-linear methods, which demands better labels of functional single-cell readouts. We, therefore, are still waiting for the "ImageNet moment" in single-cell genomics.