This is an online seminar. Registration is required.
【Team】Imperfect Information Learning Team
【Date】2025/April/7(Monday) 16:00-17:00(JST)
【Speaker】Wei Wang, RIKEN
Title: Weakly Supervised Machine Learning Revisited: Minimizing Supervision, Assumptions, and Practical Gaps
Abstract: Deep learning has achieved great success in recent years, and this success is based on the availability of large datasets with high quality annotations. However, this requirement may not be met in many real-world applications. Weakly supervised learning aims to learn an accurate model from incomplete, inexact, or inaccurate supervision. In this talk, I will discuss recent progress on this problem, including minimizing supervision information, data generation assumptions, and practical gaps. First, I will introduce a new classification setting, called confidence-difference classification, where we can learn a classifier using only unlabeled data pairs with a confidence difference indicating the differences in the probabilities of being positive. Second, I will introduce a new consistent approach to complementary-label learning that requires milder assumptions and shows the relationship between complementary-label learning and negative-unlabeled learning. Finally, I will present pitfalls in the evaluation process of the partial-label learning literature and our efforts for a more practical evaluation of this problem.
