Doctoral Thesis Proposal - Kevin Kuo
April 27, 2026 12:00PM—1:30PM
Location:
4303
-
Gates and Hillman Centers
Speaker:
KEVIN KUO,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://imkevinkuo.github.io/
Data for AI systems is becoming increasingly sparse. Within a decade, LLMs are projected to be trained on datasets the size of the total stock of public human data, while individuals and organizations are increasingly restricting access to their data due to economic and privacy concerns. Collaborative learning has the potential to fuel data-hungry AI systems by enabling access to restricted sources of data---but only if it can provide meaningful guarantees regarding data privacy, quality of service, and computational cost. This thesis studies three unique ML system architectures that offer data protection by design: (1) multi-round federated learning, (2) model merging-and-localization, and (3) proxy tuning. We leverage principles from model compression and transfer learning to improve the utility, efficiency, and privacy guarantees of these frameworks.
Thesis Committee:
Virginia Smith (Chair)
Aditi Raghunathan
Gauri Joshi
Holger Roth (NVIDIA)
Contact
Matt Stewart