Jiachen Wang (Princeton)- Fueling Responsible AI with Data Attribution
Abstract: Understanding how training data shapes model behavior is fundamental to building trustworthy AI systems. Data attribution techniques quantify the influence of individual training examples on machine learning models, providing key insights for developing data-centric algorithms (e.g., data curation) as well as addressing data-related challenges (e.g., privacy, safety, and copyright protection).
In this talk, I will present our recent advances in the foundations and practical frameworks of data attribution. First, I will introduce a general, game-theoretic data attribution framework that optimizes for stochastic learning algorithms. I will then discuss how we can efficiently conduct data attribution in the challenging setting of large-scale deep learning models (e.g., large language models). These techniques guide data quality management, explain model predictions, and boost trustworthy AI development from a data-centric perspective.
Speakers
Jiachen Wang
Jiachen (“Tianhao”) is a Ph.D. student at Princeton University, advised by Prof. Prateek Mittal. His research focuses on developing theoretical foundations and practical tools for trustworthy machine learning from a data-centric perspective. Most recently, he has been developing scalable, theoretically grounded data attribution and curation techniques for foundation models. His contributions have been recognized through multiple fellowships and oral/spotlight presentations at top AI/ML venues. He was selected as a Rising Star in Data Science in 2024.
