Transfer learning
Principled methods for combining source and target data while quantifying the statistical cost of adaptation.
Statistics · Machine Learning · Transfer Learning
Subha Maity
Assistant Professor
Department of Statistics and Actuarial Science
University of Waterloo
I develop statistical models and theory for learning reliably across populations, domains, and distribution shifts.
About
Statistical foundations for learning in changing environments
My current work focuses on transfer learning and learning under distribution shift—how to borrow information across related datasets without losing reliability in the target population.
I received my PhD in Statistics from the University of Michigan, advised by Yuekai Sun and Moulinath Banerjee. Before that, I studied mathematics and statistics at the Indian Statistical Institute, Kolkata.
Research focus
Research areas
Distribution shift
Prediction and inference when populations, labels, or strategic behavior change between training and deployment.
Reliable machine learning
Statistical perspectives on fairness, weak supervision, routing, and the behavior of modern learning systems.
Recent work
Selected publications
The Statistical Cost of Adaptation in Multi-Source Transfer Learning
A. Chakraborty and S. Maity
Transfer Learning under Group-Label Shift: A Semiparametric Exponential Tilting Approach
M. Cheng, S. Maity, Q. Tian, and P. Li
A linear adjustment-based approach to posterior drift in transfer learning
S. Maity, D. Dutta, J. Terhorst, Y. Sun, and M. Banerjee
Contact
Research and collaboration inquiries
Emailsmaity [at] uwaterloo [dot] ca
OfficeM3-4227 · University of Waterloo