알기쉬운 인공지능 튜토리얼

최승진 교수
(CROID Research & aSSIST University)

Biography
2026-현재	석좌교수, aSSIST University
2022-2025	연구소장, (주)인텔리코드
2019-2021	CTO, BARO AI & 상임고문, BARO AI Academy
2001-2019	POSTECH 컴퓨터공학과 교수
2019-2021	정보과학회 인공지능소사이어티 회장
2018	삼성전자 종합기술원 자문교수
2017-2018	삼성리서치 AI 센터 자문교수

From Belief to Guarantee: A Gentle Introduction to Conformal Bayes (60분)

A Bayesian model gives a rich language for uncertainty. It returns not only a prediction, but a predictive distribution over possible outcomes, from which we can form intervals that adapt to the data and reflect the model’s belief. The difficulty is that a Bayesian interval is calibrated to the model’s assumed world, and its advertised meaning can weaken when the model is misspecified, the data are scarce, or the training environment differs from the one in which decisions are made. Conformal prediction offers a different kind of reliability. Under exchangeability, it can wrap around almost any predictive model and deliver finite-sample coverage without assuming that the model is correct. The guarantee is not a promise that every individual interval is correct. It is a marginal coverage guarantee over repeated cases, but it is honest in a way that model-based confidence alone may not be. Conformal Bayes brings these two views together. The Bayesian model supplies a predictive distribution that captures local structure and defines what outcomes should be regarded as surprising. The conformal procedure then calibrates this Bayesian score so that the resulting prediction set has a valid coverage guarantee. When the Bayesian model is well aligned with the data, the intervals can remain sharp and adaptive. When the model is imperfect, conformal calibration provides a safeguard that the Bayesian interval alone does not. This talk introduces the idea gently, starting from the basic distinction between belief and guarantee. It explains conformal prediction, Bayesian predictive uncertainty, and how the two combine in conformal Bayes.

송경우 교수
(연세대)

Biography
2026-현재	연세대학교 응용통계학과/통계데이터사이언스학과 부교수
2023-2026	연세대학교 응용통계학과/통계데이터사이언스학과 조교수
2021-2023	서울시립대학교 인공지능학과 조교수

Can We Trust Large Language Models? Statistical, Mechanistic, and Scenario-Based Approaches to Trustworthy AI (60분)

As Large Language Models become embedded in search, writing, coding, evaluation, and decision-support systems, the question is no longer only whether they are accurate, but whether they can be trusted. This lecture explores Trustworthy AI through three complementary perspectives. The first is statistical reliability: how Conformal Inference can be used to detect hallucinations and provide calibrated confidence guarantees for LLM outputs and LLM-based evaluators. The second is mechanistic understanding: how circuit analysis and natural language autoencoders can help us interpret the internal representations and decision pathways of language models. The third is scenario-based impact measurement: how we can evaluate the real-world consequences of AI behavior under concrete deployment scenarios.

이창희 교수
(고려대)

Biography
2024-현재	고려대학교 인공지능학과 조교수
2021-2024	중앙대학교 AI학과 조교수
2019-2019	Public Health England, 방문연구원
2013-2016	한국전자통신연구원 연구원
2025-2025	IEEE Seoul Section, Young Professional Section Chair
2025-현재	한국통신학회, 학술이사
2023-현재	대학의료인공지능학회, 학술위원
2016-2021	캘리포니아 대학교, 로스앤젤리스(미국), 공학박사
2011-2013	고려대학교 전자전기컴퓨터공학과, 공학석사
2007-2011	고려대학교 전기전자전파공학부, 공학사

Interpreting with Features: A Deep Learning Approach at Global and Instance Levels (60분)

Identifying which features drive target outcomes or capture the behavior of a black-box model is essential for both scientific discovery and for revealing underlying mechanisms. Deep learning-based feature selection offers a powerful framework to achieve this. In this talk, I formulate global and instance-level feature selection as a deep learning problem and present recent advances that enhance scientific insights and model interpretability across diverse applications.

오유진 교수
(연세대)

Biography
2026-현재	연세대학교 의과대학/의생명시스템정보학과 조교수 인공지능대학원/인공지능학과 AI+X교수
2024-2026	Massachusetts General Hospital (MGH) and Harvard Medical School Postdoctoral Researcher
2019-2024	KAIST AI대학원 공학박사
2014-2019	LG전자 CTO부문 SW선임연구원
2012-2014	고려대학교 바이오의공학부 공학석사
2008-2012	고려대학교 바이오의공학부 공학사

신뢰할 수 있는 멀티모달 의료 AI 설계: 이론 중심의 고찰 (60분)

Clinical decision-making is fundamentally shaped by regional demographics and institutional protocols. However, current medical AI paradigms often rely on high-prevalence data patterns, which reinforces biases and fails to capture the nuances of diverse clinical expertise. To address this, we propose a hierarchical approach bridging engineering control theory and deep learning.
First, we introduce the distribution-aware Mixture of Experts (dMoE) as a core mechanism. Inspired by reliable control theory, dMoE is designed to adapt to non-uniform, heterogeneous data distributions. We provide a rigorous analysis of dMoE’s capacity to maintain robust performance in medical image segmentation despite distribution shifts.
Second, we extend this mechanism into the multimodal Mixture of Multicenter Experts (MoME) framework, a decentralized architecture designed to integrate specialized expertise across institutions. In this framework, Large Language Models (LLMs) are employed to encode unstructured clinical notes into rich clinical context, while the MoME modules specifically capture and aggregate institutional knowledge from diverse clinical strategies without raw data sharing. By aligning these latent clinical insights and institutional expertise with radiological features, the framework achieves a synergistic multimodal integration that significantly enhances both global generalizability and local adaptability.
We validated this approach using a multimodal target volume delineation model for prostate cancer radiotherapy. By employing few-shot training that fuses volumetric imaging with unstructured clinical notes, our model significantly outperformed state-of-the-art baselines. The performance gains were most pronounced in scenarios with high inter-center variability or data scarcity. Ultimately, MoME facilitates seamless model customization to local clinical preferences while preserving data privacy. By grounding multimodal learning in engineering control principles, this research promotes the development of reliable, scalable, and broadly generalizable medical AI.

[1] Y. Oh, P. Jin and S. Park et al., “Distribution-aware Fairness Learning in Medical Image Segmentation from A Control-Theoretic Perspective”, International Conference on Machine Learning (ICML), Top-2.6% Spotlight, 2025, https://openreview.net/pdf?id=BUONdewsBa
[2] Y. Oh and S. Park et al., “LLM-driven multimodal target volume contouring in radiation oncology,” Nature Communications, vol. 15, 9186, 2024, https://www.nature.com/articles/s41467-024-53387-y

박경덕 교수
(연세대)

Biography
2022-현재	연세대학교 응용통계학과/양자정보학과 교수
2024-현재	연세대학교 양자정보기술연구원 부원장
2021-2022	성균관대학교 나노과학기술원 연구교수
2018-2021	한국과학기술원 전기및전자공학부 연구조교수
2015-2018	한국과학기술원 물리학과 박사후연구원
2010-2015	University of Waterloo 물리학 박사
2004-2010	University of Waterloo 수리물리학 학사

Physics Meets AI: Quantum Computing and Generative Modeling (60분)

Generative AI is rapidly expanding beyond text, image, and code generation to scientific applications such as drug discovery, materials design, optimization, and quantum simulation. At its core, generative learning is the problem of modeling complex high-dimensional probability distributions and generating useful samples from them. This perspective naturally connects modern AI with statistical physics and quantum computing. In this talk, I will discuss this connection in two complementary directions. First, I will introduce how quantum computing can support generative learning through energy-based models, focusing on Boltzmann machines and their relation to quantum annealing. I will also discuss how this perspective can be connected to modern generative AI architectures. Second, I will discuss how generative learning can support quantum simulation. In particular, I will discuss how generative models can learn quantum measurement distributions and help estimate ground-state energies with fewer direct quantum-circuit measurements. Together, these examples illustrate a bidirectional relationship between generative AI and quantum computing: quantum devices as sampling resources for AI, and generative models as computational tools for quantum simulation.

주재걸 교수
(KAIST)

Biography
2025.09-현재	KAIST 김재철AI대학원 교수
2020.03-2025.08	KAIST 김재철AI대학원 부교수
2019.09-2020.02	고려대학교 인공지능학과 부교수
2015.03-2019.08	고려대학교 컴퓨터학과 조교수
2011.12-2015.02	Research Scientist, Georgia Tech
2013	Computational Science and Engineering, Georgia Tech 박사
2009	Electrical and Computer Engineering, Georgia Tech 석사
2001	서울대학교 전기공학부 학사

Recent Trends in Physical AI Research (60분)

Training AI models that can control robots poses various challenges that needs to be solved. This talk covers recent trends in physical AI research, especially addressing the issues in data acquisition, model training, and evaluation. Afterwards, I will present some of my recent research along this line, including the following:

EgoX: Egocentric Video Generation from a Single Exocentric Video, Kang et al., CVPR’26 ACG: Action Coherence Guidance for Flow-Based Vision-Language-Action Models, Park et al., ICRA’26
3D HAMSTER: Hierarchical VLAs through 3D Trajectory Guidance, Hwang et al., IROS’26 PHUMA: Physically-Grounded Humanoid Locomotion Dataset, Lee et al., under review.