Jiachen Wang (Princeton)- Fueling Responsible AI Advancement with Data Attribution

Abstract: As artificial intelligence (AI) systems expand across society, understanding how training data shapes model behavior has become fundamental to building trustworthy AI. Data attribution techniques quantify the influence of individual training samples on machine learning models, enabling us to address pressing challenges around data quality, training efficiency, copyright disputes, and interpretability.
In this talk, I will present our advances in developing theoretically rigorous yet practical data attribution methods. First, I will introduce Data Banzhaf, a data value notion derived from cooperative game theory that provides provably robust data influence estimation for any learning algorithms. While this provides a general framework, we then develop specialized techniques to analyze how data influence evolves during deep learning optimization. Through this lens, we uncover that examples from early and late training stages have an outsized impact on foundation model pretraining—insights that enable strategic data selection to reduce computational overhead while maintaining model performance.