Although deep learning models (including Transformer networks) can achieve state-of-the-art performance on computer vision and NLP tasks, gradient-boosted tree ensembles still outperform DNN for tabular data problems (which are widespread in the industry). In this paper, Borisov et al. review, categorize and evaluate existing deep learning approaches for classification and regression tasks on tabular data. In addition, they discuss methods for tabular data generation and to explain deep models for tabular data. Upon confirming that decision tree ensembles achieve higher accuracy and less training time than DNN on three benchmarks, they assert the need for better tabular preprocessing techniques; special-purpose architectures; regularization; synthetic data, and online learning to bridge this gap.