Postdoc on Tabular Foundation Models at Inria

Context: Tabular data arguably holds the most precious information in an organization. The corresponding machine learning setting, tabular learning, has recently seen impressive progress with advanced deep learning methods, in particular table foundation model –Including the TabICL model, developed at Soda–, outperforming the long-running industry standard based on gradient-boosted trees.

The progress has led to tabular learners that work very well to predict on numerical tables without tuning. Ongoing work is adding the ability to readily fit on tables non-numerical data, such as strings. Here, a promise is that blending table foundation models with large language models brings general knowledge and a form of understanding to tables.

Goals: The goal of this postdoc proposal is to push tabular learning to the next level, and to bring broader understanding of the table.