M.Sc. Bjarne Hiller, research assistant at the chair of Intelligent Data Analytics (University of Rostock), will give a presentation in the topic
"TabPFN: A Foundation Model for Tabular Data"
as part of the VAC Colloquium.
After the presentation we look forwards to discussions while enjoying coffee and cookies.
The event is open to anyone interested.
Abstract:
Large Language Models have shown capability to utilize In-Context Learning (ICL): They can simply adapt to a new task by using examples provided in the prompt, therefore allowing practitioners to avoid training a new model from scratch or fine-tuning a pre-trained model. So - what happens, if we just include a whole dataset in the prompt? Maybe even tabular datasets, which are traditionally challenging to handle using Machine Learning methods? In this presentation, we will have a look at the recent work by Hollman et al. that presented TabPFN, a foundation model for tabular datasets. We will look at the beginnings of Prior Fitted Networks (PFN), and the proposed changes to the transformer architecture that enable TabPFN to excel on tabular data. Finally, there will be a short demonstration how TabPFN can be applied in practice. Time will tell if TabPFN is just a hype or the beginning of a paradigm shift.