Speaker

Bohyung Han

Title

Beyond Ground-Truth for Discriminative Learning

Abstract

Ground-truths are often treated as unquestionable anchors for supervised learning, yet in practice they are frequently noisy, inconsistent, or incomplete. This talk challenges the assumption that ground-truths alone are sufficient, and argues for rethinking their role in discriminative learning. I will discuss how one can construct richer supervision by enhancing imperfect references—using generative priors to produce training signals that better reflect perceptual quality—and how models can reconcile heterogeneous label spaces without demanding new annotations. These examples highlight a broader agenda: progress in computer vision requires moving beyond the passive use of ground-truths toward actively shaping and expanding the supervision that drives learning.

Bio

Bohyung Han is a Professor in the Department of Electrical and Computer Engineering at Seoul National University, Korea. He was a visiting faculty researcher at Google DeepMind, Google Research, and Snap Research, and an Assistant/Associate Professor in the Department of Computer Science and Engineering at POSTECH, Korea. He received his Ph.D. from the Department of Computer Science at the University of Maryland, College Park, MD, USA, in 2005. He has served or will serve as an organizing and senior-level program committee member in major computer vision and machine learning conferences numerous times including a Program Chair in ICCV 2025 and Senior Area Chair in NeurIPS, CVPR, ICLR, and ICML. He is also an Associate Editor of TPAMI. He received the Google AI Focused Research Award in 2018. His research interests include computer vision and machine learning with an emphasis on deep learning.

Language

English · Offline