08/05/2026
By Jiali Cheng

The Kennedy College of Science, Richard A. Miner School of Computer & Information Sciences, invites you to attend a doctoral dissertation proposal defense by Jiali Cheng, titled Learning to Forget: Algorithms, Evaluation, and Mechanistic Understanding for Machine Unlearning in Artificial Intelligence.

Defense details

  • Date: Aug. 18, 2026
  • Time: 3 – 4 p.m. EDT
  • Location: Via Zoom

Committee members

  • Hadi Amiri, assistant professor, Miner School of Computer and Information Sciences, UMass Lowell (advisor)
  • Tingjian Ge, professor, Miner School of Computer and Information Sciences, UMass Lowell
  • Sijia Liu, associate professor, Michigan State University
  • Chirag Agarwal, assistant professor, University of Virginia

Abstract

Artificial intelligence (AI) models are increasingly deployed in settings where learned information may later need to be removed. A trained model may encode private user data, outdated information, harmful associations, biased patterns, copyrighted content or unsafe capabilities. Removing such information is difficult because AI models do not store knowledge as isolated records. Instead, information is spread across parameters, representations, modalities, graph structures and tool-use behaviors. Therefore, simply deleting the original data or suppressing certain outputs is often insufficient. This dissertation studies “machine unlearning” as a general framework for removing targeted information, associations or capabilities from trained AI models while preserving useful knowledge on the remaining data. This dissertation is organized into three parts:

Part I develops unlearning algorithms for multimodal models, large language models and graph models. Forgetting in these settings requires breaking specific cross-modal associations while maintaining both unimodal and multimodal representation quality; removing the influence of nodes, edges or subgraphs in the context of message passing without damaging the model’s performance on the remaining graph; and unlearning not only a data point but also a learned behavior or functional capability.

Part II develops a cross-modal and cross-task evaluation benchmark for machine unlearning algorithms. This benchmark introduces unlearning settings that have been largely unexplored in prior work, including biomedical relation extraction, visual reasoning, speech processing and video understanding. An effective unlearning algorithm should not only forget targeted information but also preserve performance on retained knowledge, remain robust under out-of-distribution settings and protect private information against adversarial attacks. The benchmark demonstrates that existing unlearning algorithms often struggle to satisfy these goals simultaneously.

Part III analyzes unlearning from the perspectives of loss landscape and mechanistic interpretability. These studies introduce mode connectivity to the unlearning setting and show that independently unlearned models can converge to the same low-loss region. They also show that merging multiple unlearned models strengthens models’ robustness to relearning attacks. Using circuit-based mechanistic analysis, these studies explain why some samples can be easily removed from a model, while others are more difficult to unlearn.

These contributions form a connected view of machine unlearning. Part I develops algorithms for removing information from different types of AI models. Part II evaluates these algorithms across diverse tasks and reveals where they succeed or fail. Part III explains the underlying mechanisms behind these successes and failures, which can guide the design of stronger algorithms and more reliable evaluation methods. Overall, this dissertation establishes machine unlearning as a system-level capability for trustworthy AI, where forgetting must be designed, optimized, evaluated and interpreted across the full life cycle of AI systems.