Invited Talks

장형진 교수
(버밍엄대)

Biography
2022-현재	Associate Professor @ University of Birmingham
2025-현재	Affiliated Researcher @ The Institute of Data and AI (IDAI)
2021-2023	Turing Fellow @ The Alan Turing Institute
2018-2022	Assistant Professor @ University of Birmingham
2013-2017	Post-doc Researcher @ Imperial College London

When AI Sees People: Human-Centred Vision and Interaction (45분)

AI is created by humans and for humans. In this talk, I will introduce our recent research on human-centred visual learning — building AI systems that understand people better by focusing on how we move, look, and interact with the world. Key visual cues like body pose, hand movements, gaze direction, and object interaction provide important insights into human actions and intentions. Our team has been working on combining these visual signals with language models to help AI systems understand context more effectively and interact more naturally with people. I will share our latest work on integrating body motion, hand gestures, and gaze with language to build smarter and more responsive systems. We also explore how different types of information — such as audio, text, and images — can work together to improve tasks like estimating hand-object shapes or generating realistic face and gaze images. This multimodal approach makes AI more accurate, flexible, and easier to communicate with. Finally, I will highlight recent trends in computer vision and discuss how these ideas can shape the future of human-AI interaction. These technologies have the potential to change the way AI sees and connects with people in everyday life.

정다샘 교수
(서강대)

Biography
2021-현재	서강대학교 아트&테크놀로지학과 조교수
2024-현재	ISMIR 2025 공동조직위원장
2025-현재	한국음악정보학회 부회장
2024-현재	한국음악지각인지학회 부회장
2025-현재	판소리학회 섭외이사
2020-2021	SK Telecom T-Brain 연구원
2015-2020	KAIST 문화기술대학원 공학박사
2013-2015	KAIST 문화기술대학원 공학석사
2008-2013	KAIST 기계공학과 공학사

Deep Learning for Diverse Music Data (45분)

Music exists in various formats. It can be in the form of a musical score (either scanned or in a machine-readable format like MusicXML), or as data with more performance-related aspects, such as MIDI or audio. Furthermore, different music genres often use different forms of representation. For example, in the case of Korean court music, music exists in Jeongganbo. This talk introduces two studies concerning the handling of different formats of music data. The first study focuses on reviving 15th-century Korean court music. This involves two key contributions: first, constructing the first machine-readable Jeongganbo dataset using optical music recognition, and second, proposing a novel encoding method that imitates Jeongganbo notation to represent melodies. The second study introduces a unified framework for cross-modal translation among various music formats, including scanned score images, machine-readable scores, MIDI performance data, and audio recordings.

최성준 교수
(고려대)

Biography
2025.01 - 현재	Principal Scientist at RLWRLD
2020.09 - 현재	Assistant Professor at Korea University
2018.06 - 2020.08	Postdoctoral Associate at Disney Research
2018.02 - 2018.05	Research Scientist at Kakao Brain
2012.09 - 2018.02	Ph.D. in EECS, Seoul National University

Towards Human-Centered Robotics: Coexistence of Human and Robots (45분)

Recent advancements in large language models and vision-language models have opened up new possibilities for human-centered robotics. This presentation aims to explore the potential of these cutting-edge technologies in enhancing human-robot interaction and enabling robots to better serve human needs. Firstly, we will see the rapid development of humanoid robots and their potential to replace human labor in various domains. Secondly, we will see the integration of artificial intelligence techniques with these humanoid robots to facilitate their deployment in everyday life scenarios. Furthermore, I will showcase our laboratory's ongoing research projects including utilizing Vision Language Models (VLMs) for human-robot interaction, combining VLMs with simulators, and developing robot agents with distinct personas.

김선 교수
(서울대)

Biography
2011-현재	서울대 컴퓨터공학과 교수
2022-현재	서울대 생명과학부 겸임교수
2024-현재	서울대병원 AI Healthcare 교수
2021-현재	(주) 아이겐드럭 CEO/CTO
2024-2025	국가인공지능위원회 인재/인프라 분과 위원장
2022-2024	(재) 목암생명과학연구소, 소장
2017-2019	한국정보과학회 인공지능소사이어티, 회장
2009-2011	Indiana University School of Informatics and Computing, 학과장
2001-2011	Indiana University School of Informatics and Computing, 조/부교수
1998-2001	DuPont Central Research, Senior Computer Scientist
1997-1998	UIUC Keck Center Bioinformatics Unit, Director

Navigating Chemical Space with Deep Learning Technologies (45분)

Drug development is a long and complex discovery process. Now advanced deep learning technologies have begun to realize AI-based drug discovery with a pronounced example of protein structure prediction (2024 Novel prize). In this talk, I will share our recent works on navigating chemical space with deep learning technologies which occurs in early stages of drug discovery. First, we show how drug-likeness can be re-defined with deep learning technologies (ISMB 2025, ICML 2025). Then, we will share our works on predicting drug target affinity prediction (ICLR 2025, ISMB 2025).

이종민 교수
(연세대)

Biography
2025-현재	연세대학교 인공지능학과 조교수
2022-2025	UC Berkeley, 박사후 연구원
2021	Google DeepMind, Research Scientist Intern
2022	KAIST 전산학 박사
2017	KAIST 전산학 석사
2014	서울대학교 컴퓨터공학 학사

RL Meets Convex Optimization (45분)

Reinforcement learning (RL) has achieved impressive results in domains like games and robotics, but its reliance on large amounts of on-policy interaction limits its real-world applicability. In this talk, I will present a unified framework that enables data-efficient policy learning from arbitrary off-policy data, built on top of convex optimization. By directly optimizing the stationary distribution of policies, rather than relying on unstable off-policy actor-critic methods, our approach achieves stable and principled learning across diverse settings such as reward-maximizing RL, constrained RL, imitation learning, state entropy maximization, and more. This perspective simplifies algorithm design and demonstrates that the objective for many complex sequential decision-making problems can be reduced to solving a single convex optimization.

윤세영 교수
(KAIST)

Biography
2017	KAIST 김재철AI대학원 부교수
2016-2017	Los Alamos National Lab. (미국) 박사후연구원
2015-2016	Microsoft Research, Cambridge (영국) 방문연구원
2014-2015	Microsoft Research-INRIA (프랑스) 박사후연구원
2013-2014	KTH (스웨덴) 박사후연구원
2012	KAIST 전기및전자공학 박사
2006	KAIST 전기및전자공학 학사

Training for Efficient LLM Inference (45분)

Recently, large language models (LLMs) have significantly impacted our lives. Services such as ChatGPT, Gemini, and Claude are powered by LLMs and offer AI capabilities for a wide range of applications. However, these models typically require substantial computational resources and are primarily deployed in the cloud, rather than on-device. For reasons such as privacy concerns and offline use cases, on-device deployment is becoming increasingly important. This highlights the need for methods that can reduce inference costs without compromising performance. In this talk, I will present several training strategies aimed at achieving this goal, including knowledge distillation, block decoding, and weight reuse techniques.

신원용 교수
(연세대)

Biography
2019-현재	연세대학교 수학계산학부(계산과학공학) 부교수/교수
2024-현재	연세대학교 배터리공학과 겸임교수
2022-현재	POSTECH 인공지능대학원 겸임교수
2021-현재	㈜카이로스랩 공동창업자
2024-현재	㈜프리딕션 공동창업자
2012-2019	단국대학교 컴퓨터학과 조교수/부교수
2009-2012	Harvard University, 박사후연구원/Research Associate
2008	KAIST 전자전산학과 공학박사

초고속 추천용 그래프 필터링 방법 (45분)

Graph convolutional network (GCN) has received considerable attention as a neural network structure that learns the optimal weight parameters by combining node attribute information with graph connection information for various graph downstream tasks. However, GCN-based recommender systems have the disadvantage in that they take too long to learn to reflect the large-scale interaction data and the user's preferences that change in real time. In this talk, I present a new methodology that performs recommendation based on pre-defined low-pass filters without expensive training processes. Built upon the pioneering work, dubbed GF-CF, we newly design Turbo-CF, a more cost-effective graph filtering method that does not necessitate costly matrix decomposition. Through comprehensive experiments, I demonstrate the effectiveness of Turbo-CF in terms of runtime efficiency and recommendation accuracy.