Yuan He

My PhD research focuses on developing privacy-preserving and interpretable machine learning methods. A central part of my research is differential privacy (DP), which provides formal guarantees on the privacy of individual training examples. I investigate how differential privacy can be applied to modern machine learning systems while maintaining good performance. One of my current research directions focuses on differentially private continual learning, where models learn from a sequence of tasks while protecting training data and mitigating catastrophic forgetting. In this setting, I explore interpretable approaches based on class prototypes, as well as methods for preserving previously learned representation structures. Beyond continual learning, I am interested in extending privacy-preserving and interpretable learning to multimodal and vision-language models, particularly when models are adapted using data containing personal or sensitive information. Overall, my PhD aims to develop machine learning methods that combine strong privacy guarantees with interpretability and practical utility.