MEETUP LABELIA #4 : Amortizing Label Shift in Tabular Foundation Models
S’inscrire à l’événement
Program:
- Talk by Mariana Vargas Vieyra, ML and AI researcher at Rhizome Labs, on amortizing label shift in Tabular Foundation Models
- Open discussion on responsible and trustworthy AI, and on how Labelia can keep pace with a fast-evolving field
After a deep dive into causality and explainability with Marianne Clausel last time, this edition continues with another technical session. Mariana Vargas Vieyra, an ML and AI researcher at Rhizome Labs, will present ongoing work with Clément Poiret on amortizing label shift in Tabular Foundation Models (TFMs), a question that touches on a key aspect of trustworthy AI: how reliable models remain when data distributions change.
Thanks to Prior Fitted Networks, TFMs can be pre-trained entirely on synthetic data and applied to new problems through In-Context Learning, without retraining. They have proven successful across a broad variety of real-world tasks, but they assume that train and test distributions are fixed, which is often unrealistic. In particular, they are vulnerable to label shift, when class proportions change between training and test time.
Mariana will introduce ShiftICL, the first TFM that learns to adapt to label shift during pre-training. It handles previously unseen shifts without retraining or access to test labels and, across 209 OpenML datasets, substantially improves accuracy and calibration over state-of-the-art methods, with gains growing as the shift increases.
Rhizome Labs is developing an adaptive AI engine that maintains model performance as conditions change. Lancelot Convert, CEO of Rhizome Labs, will also join us for the evening.
Labelia is the reference label for responsible and trustworthy AI, built on three pillars: an open-source framework, a self-assessment platform and a label. Behind it is a committed community of experts from diverse backgrounds, who help shape the framework so that it stays relevant in a fast-evolving field. Want to join them? Contributions of all kinds are welcome, and this meetup is a great place to start. Whether you are already involved or simply curious about trustworthy AI, come and share your perspective!