As machine learning (ML) and AI evolve, so do concerns around data privacy. Department of Statistical Sciences (DOSS) Professor and Vector Research Director Daniel Roy and collaborators delved into this issue in their award-winning paper, "Information Complexity of Stochastic Convex Optimization: Applications to Generalization and Memorization".
Presented at the International Conference on Machine Learning (ICML) in Vienna, Austria, the paper was among only 10 awarded Best Paper out of over 10,000 submissions. Their research reveals that optimal learning algorithms must memorize a constant fraction of their training data, raising significant privacy concerns. The findings emphasize the need for methods to “unlearn” data, which has critical implications for protecting sensitive information in ML systems.
The paper addresses stochastic convex optimization, a mathematical model covering many types of regression and basic neural networks, with implications that may extend to modern neural networks. These insights shed light on the balance between efficient learning and data privacy.
Learn more about this award-winning paper.