AI & Fundamentals

Highly Accurate Predictions on Small Data with the Tabular Foundation Model TabPFN - Frank Hutter, Professor, University of Freiburg

Abstract:

Tabular data, spreadsheets organized in rows and columns, is ubiquitous in many fields, prominently including medicine. The fundamental prediction task of filling in a label column based on the rest of the columns (the so-called features), is essential for predictive diagnostics, biomedical risk models, and drug discovery. Yet, in contrast to the deep learning revolution for text and images, the traditional machine learning methods used for such tabular data have made almost no progress for a decade. In this talk, I discuss TabPFN, the first foundation model for tabular data to dramatically improve predictive performance over traditional methods. For datasets with up to 10000 data points and 500 features, TabPFN yields better performance in seconds than any previous method in hours. Based on this, we expect TabPFN to become the new default method for small tabular data prediction. 

Bio:

Frank Hutter is a Hector-Endowed Fellow and PI at the ELLIS Institute Tübingen, as well as Full Professor for Machine Learning at the University of Freiburg (Germany). He is a Fellow of EurAI and ELLIS, the director of the ELLIS unit Freiburg and the recipient of 3 ERC grants. Frank was at UBC for 9 years, as a visiting graduate student, PhD student and postdoc. He is best known for his research on automated machine learning (AutoML). He co-authored the first book on AutoML, won the first two AutoML challenges with his Auto-sklearn team, is co-teaching the first MOOC on AutoML, co-organized 15 AutoML-related workshops at ICML, NeurIPS and ICLR, and founded the AutoML conference as general chair in 2022 and 2023. In recent years, his focus has been on tabular foundation models; he just co-founded PriorLabs, the world’s first startup on tabular foundation models.