Although applications of deep learning to images and natural language captivate the public’s attention, many (if not most) companies today are applying neural networks to structured data. In this paper, Somepalli et al. adopt the popular Transformer-based modeling approach to develop a new deep learning architecture for tabular data problems, SAINT. SAINT projects all tabular data features (which are subsequently tokenized) into a combined dense vector space. In addition to leveraging self-attention, SAINT also uses “intersample attention” to relate a data sample to other similar rows in the table. The authors demonstrate that SAINT can outperform popular data science methods like boosted trees on several benchmark tasks.