Efficient Tabular Foundation Models and Learned Table Embeddings

Tabular foundation models have recently revolutionized AI on structured data, offering an unprecedented level of predictive performance compared to classical Machine Learning methods. In this talk, Josif will present his research targeting two aspects of tabular data representation: efficient inference for tabular foundation models and self-supervised table embeddings.

Concerning the first aspect, tabular foundation models are lazy learners that incur high inference costs because they predict a test instance by capturing its interactions with the full support of the training data points. To this end, Josif will present his most recent research on scaling up tabular foundation models by compressing the training set into a latent representation. For the aspect of table embeddings, Josif will introduce his research on modeling tables by training dataset meta-features through synthetic self-supervised learning.

Speaker: Prof. Dr. Josif Grabocka has been a full professor of Machine Learning at UTN since December 2023. Before that, he was an assistant professor (W1) of Representation Learning at the University of Freiburg (2019–2023), after receiving his Ph.D. degree from the University of Hildesheim in 2016. Prof. Grabocka’s research interests focus on the optimization of machine and deep learning pipelines from the perspective of hyperparameter optimization. In addition, he is interested in optimal control and reinforcement learning, learning from tabular data, as well as multi-objective optimization.

Prof. Grabocka is an Area Chair at the International Conference on Learning Representations (ICLR) and a Senior Area Chair at the International Conference on Automated Machine Learning (AutoML). Since 2025, Prof. Grabocka has been a member of the ELLIS Society. His research has been published in prestigious international venues in the field of machine and deep learning, such as ICLR, NeurIPS, SIGKDD, etc.

At UTN, he has been the Coordinator of the M.Sc. program on AI and Robotics (AIR) since 2024. Prof. Dr. Grabocka teaches courses in the AIR program on Deep Learning and Large Language Models.

‍