Learning invariant models for out-of-domain generalization: good news and bad news
Date: February,14,2024 Start Time: 10:30 - 11:30
Location: 1061, Meyer Building
Lecturer: Uri Shalit
|I will present work showing how a robust notion of model calibration ties into the idea of invariant learning, and leads to models that can generalize out-of-domain (OOD) in both theory and practice. I will then show how practical difficulties with optimizing the above models lead us to a new result with ramifications for OOD generalization, fairness and robustness: We prove how “benign overfitting”, where deep models interpolate the training set yet generalize well, can be fundamentally at odds with learning invariant models.
|Uri Shalit is an associate professor at the Technion – Israel Institute of Technology. Previously, Uri was a postdoctoral researcher in Prof. David Sontag’s Clinical Machine Learning Lab at NYU and then MIT.