Slide Decks
Courses, talks, and showcases.
Web-based slide decks for my courses, external talks, and meetings. Each deck is its own page — scroll, or use ↓ / Space to advance.
Courses
Talks & Seminars
- Talk · Sep 17, 2026
‘비슷함’을 미래로 정의하기
금융 R&D 포럼 · 금융보안원 여의도 교육센터
FASCL(Future-Aligned Soft Contrastive Learning) 발표. 미래 수익률을 정확히 맞히는 대신, 앞으로 함께 움직일 자산끼리 가깝게 두는 표현 학습. Video representation learning의 경험에서 출발해 factor model, soft target, 전체 결과까지.
- Guest Lecture · Sep 11, 2026
Video Foundation Models and World Models
DGIST HSS118 · Future Literacy for the Age of Physical AI
A guest lecture for first-year students. Why video is not a stack of images (motion lives between frames), what it costs today's models to see motion, how world models predict what happens next and where generated worlds crack (the Nether challenge), and one thread through it all: trivial for us, expensive for machines. Keep proposing new directions.
- Talk · Aug 20, 2026
Not All Prediction Targets Keep Training-Free Diffusion Guidance on the Manifold
The 2nd UoM-SeoulTech International Joint Workshop · University of Miyazaki
A 15-minute invited talk on our ECCV 2026 paper. Starts from a gentle introduction to generative models and diffusion, then builds up to training-free guidance, its weak link (recovering the clean image from predicted noise), and why x-prediction (JiT) keeps guidance on the manifold. Written for an audience new to AI.
- Talk · Aug 1, 2026
잘하는 공학자에서, 믿을 수 있는 공학자로
YEHS X TI:um 공학 리더십 세미나 · 서울대학교
실력이 커질수록 공학도가 마주하게 되는 질문들. 슬견설과 공리계에서 리워드 해킹과 게임이론까지, 납득과 신뢰와 본질에 대하여.
- Seminar · Jul 5, 2026
ALOHA / ACT — Fine-Grained Bimanual Manipulation
ALOHA/ACT (RSS 2023) 논문 세미나. 오차 예산을 하드웨어로 사는 고전 스택에서 고주파 픽셀 closed-loop의 학습된 반응성으로. action chunking · CVAE · temporal ensemble의 '왜'를 따라가는 발표.
- Seminar · May 30, 2026
World Models
AI Seminar
World Model 분야 1시간 세미나. 상상으로 행동하는 기계(Dreamer)에서, 세계를 생성하는 기계(Sora·Genie·Cosmos), 그리고 로봇 frontier까지의 흐름.
- Meeting · May 27, 2026
Computer Vision for Videos — studying motion
A ~10-min introduction: who I am, why motion is the real bottleneck in video understanding, and three papers that trace the thread from frame interpolation to world models.