Is Deep Learning on Tabular Data Enough? An Assessment