NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction

NVIDIA's Kumo Tabular predicts categories or numeric values for new table rows using labeled rows supplied alongside them. The release article describes a model pretrained on artificial tables that can make these predictions in a single forward pass, without training a new model for each dataset. It examines values within columns, relationships across a row and similarities to the labeled examples. Because the labeled rows never attend to the rows being predicted, their computed context can be reused for follow-up predictions. This offers a reusable prediction engine for tables, while still requiring relevant labeled examples from the task at hand.