세미나안내
Learning What Matters: Efficient and Reliable Multimodal Intelligence
- 등록일2026.10.12
- 조회수33
-
세미나 일정2026.10.14 수
-
연사윤상두 소장(NAVER AI Lab)
[Abstract]
Building efficient and reliable multimodal intelligence requires choices about what systems learn, remember, and compute, and how they use that information. In this talk, I will present research on selective memory, adaptive inference, safety-aware learning, and predictive world models. These studies address how models retain useful context, allocate computation, condition the use of dual-use knowledge, and learn from diverse visual experience and demonstrations.They highlight a common research question: what information should a model retain, use, and predict to support its intended behavior? I will conclude with a research agenda on how these design choices extend to longer histories, new tasks, and changing environments.
[Biography]
Sangdoo Yun leads NAVER’s Future AI Center and NAVER AI Lab and is an adjunct professor at the Artificial Intelligence Institute of Seoul National University and GSAI of POSTECH. His research focuses on building efficient, reliable, and capable multimodal foundation models for real-world AI applications. At NAVER, he has led the development of machine learning methods spanning model architecture, training and optimization, robustness, multimodal learning, and large-scale AI systems. He is best known for co-inventing CutMix and has contributed to advances in model architecture, optimization, robustness, and multimodal learning. He received his B.S., M.S., and Ph.D. degrees from Seoul National University in 2010, 2013, and 2017, respectively, with his graduate research focused on computer vision. He regularly serves as an Area Chair for leading AI conferences, including NeurIPS, CVPR, ICLR, ECCV, ICCV, and ICML.
