Loading...

학술행사

Korean AI Association

  >   학술행사   >   국내학술대회

국내학술대회

튜토리얼
 
 
강민석 교수(서울대학교 데이터사이언스대학원)
 
Title:  인간 중심 AI 평가: 벤치마크 점수가 말해주지 않는 것들
 
Abs: 
생성형 AI 모델을 평가할 때 흔히 벤치마크 점수와 리더보드 순위에 주목한다. 하지만 그것만으로 어떤 AI가 '좋은' AI인지 판단할 수 있을까? 본 튜토리얼에서는 단일 순위, 집계 점수, 고정된 테스트셋에 의존하는 기존 벤치마크 중심 평가의 한계를 짚어본다. 그리고 이러한 문제를 다루기 위한 접근으로, 사람이 평가 과정에 직접 참여하도록 돕는 인간 중심의 인터랙티브 도구와 인간-AI 협업 워크플로우를 소개한다. 이를 통해 모델을 개발 또는 배포하는 사람들이 주어진 점수에만 의존하지 않고, 모델의 문제점을 직접 발견하고 자신의 필요에 맞게 평가 기준을 만들어 가는 방법을 살펴본다. 나아가, AI를 인간의 의도와 가치에 맞게 정렬(alignment)한다는 것의 의미를 생각해보고, 인간 중심 AI 평가와 관련한 최근 연구 쟁점과 향후 방향을 논의한다.
 
Bio:
Minsuk Kahng is an Assistant Professor at Seoul National University in the Graduate School of Data Science and SNU College. His research bridges responsible AI, human-computer interaction, and data visualization, with a focus on building interactive tools and human-AI collaboration workflows that support the responsible development of AI systems. Before joining Seoul National University, he was an Assistant Professor at Yonsei University, a Senior Research Scientist at Google DeepMind in the People + AI Research team, and an Assistant Professor at Oregon State University. He received his Ph.D. in Computer Science from Georgia Tech with a Dissertation Award, and his Master's and bachelor's degrees from Seoul National University.

 
 
최종현 교수(서울대학교 전기·정보공학부)
 
Title:  Challenges in Learning RFM and World Models
 
Bio:
Jonghyun Choi received the B.S. and M.S. degrees in electrical engineering and computer science from Seoul National University, Seoul, South Korea in 2003 and 2008 respectively. He received a Ph.D. degree from University of Maryland, College Park in 2015, under the supervision of Prof. Larry S. Davis. He is currently an associate professor at Seoul National University, Seoul, South Korea. He was an associate professor at Yonsei University (2022-2024), an assistant professor at GIST (2018-2022), a research scientist at Allen Institute for Artificial Intelligence (AI2), Seattle, WA (2016-2018), a senior researcher at Comcast Applied AI Research, Washington, DC (2015-2016). He serves as an area chair at NeurIPS, ICLR, CVPR, ICCV, ECCV, WACV, AAAI (senior PC), ICRA (associate editor), and an associate editor in IEEE Transactions on PAMI and IJCV. His research interest includes visual understanding for edge devices and household robots and visual recognition of images and videos under computational and data efficiency constraints.
 

 
 
김재형 교수(연세대학교 인공지능학과)
 
Title:  Aligning LLMs with Individuals and Humanity: From Universal to Personalized Alignment 
 
Abs: 
Aligning large language models (LLMs) with human preferences and values is essential to making them both helpful and harmless. This tutorial will first introduce the fundamentals of LLM alignment and review how the field has evolved in recent years. This tutorial will then turn to LLM personalization, which can be viewed as the most fine-grained form of alignment, and cover its foundations and recent paradigms. Finally, this tutorial will conclude by discussing promising future directions for alignment and personalization.
 
Bio:
Jaehyung Kim received the B.S. degree in mathematical science from KAIST, in 2017, and the M.Sc. and Ph.D. degrees in electrical engineering from KAIST, in 2019 and 2014, respectively. From Jan. 2024 to Aug. 2024, he worked as a post-doc at CMU. In Sep. 2024, he then joined a department of artificial intelligence at Yonsei University. His recent research has been featured in oral and spotlight presentations at top-tier conferences in ML and NLP, including ICLR, ICML, NeurIPS, ACL, and EMNLP.

 
주한별 교수(서울대학교 컴퓨터공학부)
 
Title:  Beyond Plausible: Physically Valid Human-Object Interaction for Robot World Models​
 
Abs: 
Equipping AI and robotic systems with the ability to understand human behavior is essential for enabling them to assist people across a wide range of everyday applications. This need is more pressing than ever: the heaviest consumers of such knowledge are no longer perception systems alone, but robot policies and world models that must learn to act in the physical world. Yet the high-quality 3D human-object interaction data required to learn this knowledge remains extremely scarce, and what generative models produce today is often plausible but not physically valid.

In this tutorial, I will present our lab's efforts to scale and enrich human-object interaction data, with the ultimate goal of teaching robots to interact with the world like humans.

I will first introduce our multi-year effort in building multi-camera capture systems, from Panoptic Studio to ParaHome, and most recently RoboHome, a new system designed to capture human-object interactions in natural home environments. Next, I will present a complementary direction: learning everyday interactions and affordances from generative image and video models, which offer a scalable source of human behavior priors that capture systems alone cannot reach. Finally, I will discuss the missing pieces that vision and image models cannot provide, including physics, contact, and the gap across diverse embodiments, together with our recent efforts to fill them with robot world models.
 
Bio: 
Hanbyul Joo is an associate professor at Seoul National University (SNU) in the Department of Computer Science and Engineering. Before joining SNU, Hanbyul was a Research Scientist at Facebook AI Research (FAIR), Menlo Park. Hanbyul received his PhD from the Robotics Institute at Carnegie Mellon University. He is a recipient of the Samsung Scholarship, the Okawa Foundation Research Grant, and the Best Student Paper Award at CVPR 2018.