2026 Theses Doctoral

Proactive Tracking and Adaptation of Model Derivatives for Reliable Machine Learning Systems

Hao, Wei

New machine learning models are often derived from existing ones (e.g., through transfer learning, continuous learning or quantization), forming an ecosystem where models are related to each other and can share structure or even parameter values. While managing such a large ecosystem of model derivatives is beneficial, it also introduces new operational challenges in debugging and updating deployed systems. For example, deploying models across different data distribution or heterogeneous hardware platforms requires multiple adapted variants, complicating debugging and coordinated updates even when models differ only slightly. Existing approaches largely treat models as independent artifacts, limiting their ability to manage such dependencies at scale.

This dissertation argues that treating model derivatives as first-class, interdependent entities enables systematic and scalable solutions to reliability challenges in modern machine learning systems. Guided by this thesis, I develop three complementary systems. Nazar enables proactive, root-cause analysis and coordinated updating across fleets of deployed models. DIVA demonstrates the necessity of differential testing because vulnerabilities may arise in model derivatives due to the adaptation process. MGit introduces explicit lineage tracking to capture complex model derivation relationships, making it easier to store, test, update, and collaborate on diversely related models, accommodating complex derivation pathways beyond a linear model creation timeline. Together, these contributions establish a unified framework for proactive tracking, diagnosis, and adaptation of model derivatives for reliable machine learning systems.