패턴인식/기계학습 세션

황성재 교수
(연세대)

Biography
2022-현재	연세대학교 인공지능융합대학 인공지능학과 부교수
2019-2022	University of Pittsburgh, Computer Science, 조교수
2014-2019	University of Wisconsin-Madison, Computer Science, PhD

From Medical Images to Multimodal Medical AI (60분)

Medical imaging AI is expanding beyond conventional prediction tasks such as classification, segmentation, and reconstruction. In this talk, I will discuss how generative models and multimodal vision-language models are reshaping medical imaging research through image synthesis, harmonization, visual grounding, and image-text understanding. I will present our recent work in medical image analysis and generative modeling, and briefly discuss future opportunities for reasoning-oriented AI systems in medical imaging.

이주용 교수
(서울대)

Biography
2024-현재	서울대학교 약학대학 및 융합과학기술대학원 부교수
2022-2024	서울대학교 약학대학 및 융합과학기술대학원 조교수
2017-2022	강원대학교 화학생화학부 조교수
2012-2017	미국 국립보건원(NHLBI/NIH) 방문연구원
2011-2012	고등과학원 계산과학부 박사후연구원
2011	서울대학교 화학과 이학박사
2007	서울대학교 화학과 이학석사
2005	서울대학교 화학·물리학 학사
2022-현재	Arontier Co. 최고기술책임자(CTO)
연구분야	AI 기반 신약 개발, 분자 동역학, 단백질 디자인, 단백질 구조 예측
주요논문	Nat. Commun., Adv. Sci., J. Chem. Inf. Model. 등 주요 국제학술지 논문 다수

Application of Generative AIs for Drug Discovery (60분)

The integration of artificial intelligence (AI) and molecular dynamics (MD) simulations is rapidly transforming the drug discovery process. Recent advancements demonstrate that state-of-the-art computational techniques are accelerating the discovery of drug candidates across various modalities, including small molecules, peptides, and antibodies. This presentation will explore recent progress in generative biomolecular modeling AI models applied to drug screening, candidate generation, with several case studies illustrating their practical applications. First, the new generative AI methods for protein design, protein sequence generation methods. The new methods show superior performance to existing inverse folding methods, especially for antibody CDR loop generation. Second, we will discuss the development of novel peptides for GLP1R activation. Lastly, we will discuss the development of a novel artificial antigen-based RSV vaccine candidate using protein generative models. In all three cases, the success rate for identifying novel candidates was significantly higher compared to traditional high-throughput screening approaches, underscoring the practical advantages of AI-driven strategies in modern drug discovery.

최동희 교수
(부산대)

Biography
2025-현재	부산대학교 정보컴퓨터공학부 조교수
2023-2025	Imperial College London - Research Associate
2022-2023	Sony Research - Research Intern
2018-2019	LYZE, Inc 공동창업자 & Research Lead
2016-2019	Kono Labs 데이터 사이언티스트
2014-2016	Opinion8 공동창업자 & CIO
2023	고려대학교 컴퓨터학과 박사
2014	고려대학교 바이오협동과정 석사
2012	고려대학교 컴퓨터통신공학부 학사
2025-현재	한국정보과학회 데이터소사이어티 이사

Relation Extraction for Diet, Non-Communicable Disease and Biomarker Association (60분)

Diet plays a critical role in human health, with growing evidence linking dietary habits to disease outcomes. However, extracting structured dietary knowledge from biomedical literature remains challenging due to the lack of dedicated relation extraction datasets. To address this gap, we introduce RECoDe, a novel relation extraction (RE) dataset designed specifically for diet, disease, and related biomedical entities. RECoDe captures a diverse set of relation types, including a broad spectrum of positive association patterns and explicit negative examples, with over 5,000 human-annotated instances validated by up to five independent annotators. Furthermore, we benchmark various natural language processing (NLP) RE models, including BERT-based architectures and enhanced prompting techniques with locally deployed large language models (LLMs) to improve classification performance on underrepresented relation types. The best performing model was gpt-oss-20B, a locally-deployed open-weight LLM, achieving an F1-score of 64% (macro) for multi-class classification and 92% for binary classification using a hierarchical prompting strategy with a separate reflection step built in. To demonstrate the practical utility of RECoDe, we introduce the Contextual Co-occurrence Summarisation (CoCoS) framework, which aggregates sentence-level relation extractions into document-level summaries and further integrates evidence across multiple documents. CoCoS produces effect estimates consistent with established dietary knowledge, demonstrating its validity as a general framework for systematic evidence synthesis.

김범준 교수
(KAIST)

Biography

Beomjoon Kim is an Associate Professor at the KAIST Graduate School of AI. He holds a PhD from MIT CSAIL, a Master's degree from McGill University, and an undergraduate degree from the University of Waterloo, all in computer science. His research goal is to enable robots to operate in complex unstructured environments by integrating learning, reasoning, and perception. He has won the ICRA Best Paper Award in 2024, the Google Research Scholar Award in 2023, and the ICRA Best Cognitive Robotics Award in 2018. He is not a robot.

Generalizing over objects and environments with fewer parameters for robot manipulation (60분)

The current idea in vogue is big model, big data, and end-to-end training for developing a general-purpose robot. But here is the problem with this approach: it consumes too much power. For instance, LLAMA 8 billion model uses 250-300 watts just to make a single inference. And that’s for a language model which only has to process a discrete set of symbols. We can only expect the power requirement would be larger for robotics, which has to process a continuous stream of high dimensional sensory data to output a sequence of continuous actions. In contrast, humans on average use only 20 watts of power. This tells us that there is something wrong with how we are building our AI models. In this talk, I will talk about our lab's recent effort to discover useful inductive biases for robot manipulation so that we can do more with less data and smaller models, much like what CNNs did for images.

임성빈 교수
(고려대)

Biography
2023-현재	고려대학교 통계학과 부교수
2020-2023	UNIST 인공지능대학원/산업공학과 조교수
2018-2019	카카오브레인 연구원
2016-2017	삼성화재 선임
2010-2016	고려대학교 이과대학 (수학/해석학) 이학박사
2005-2010	고려대학교 정경대학 정치외교학과
2023-현재	LG AI 연구원 자문

Generative AI for Causal Reasoning (60분)

Causal discovery remains challenging in high-dimensional data, where combinatorial search over graphs becomes intractable. This talk presents two complementary works. First, we show that LLMs can reason about causal relationships from semantic variable descriptions and we integrate LLM-derived priors into data-driven discovery algorithms, letting observational evidence suppress false discoveries. Second, we present a functional diffusion model for causal ordering whose statistical estimates can be coupled with prior knowledges to control LLM-based causal reasoning without fine-tuning or prompt engineering.

김은솔 교수
(한양대)

Biography
2025-현재	인공지능소사이어티 학술부회장
2024-현재	재단법인 브라이언임팩트 사내이사
2021-현재	한양대학교 컴퓨터소프트웨어학부 조교수
2018-2021	카카오브레인 비디오이해팀 팀장
2018	서울대학교 전기컴퓨터공학부 박사
2010	서울대학교 전기컴퓨터공학부 학사

Toward Dynamics-Preserving Representations for World Models and Scientific Foundation Models (60분)

Recent advances in representation learning have largely focused on preserving semantic information for downstream tasks such as classification, retrieval, or reconstruction. While these objectives have led to remarkable progress across vision and language, it remains unclear whether they are sufficient for modeling long-horizon dynamics in scientific and physical systems. In this talk, I will discuss a different perspective: a good representation for world models should preserve the underlying system dynamics rather than merely reconstruct observations. Motivated by recent developments in identifiable representation learning and latent dynamical system identification, I will explore the hypothesis that long-horizon generation quality is fundamentally determined by the geometry of the latent state space. This viewpoint naturally connects representation learning, system identification, reaction coordinates in molecular dynamics, and world models. Finally, I will present several ongoing research directions toward learning dynamics-preserving latent representations that may provide a common foundation for video generation, protein conformational dynamics, weather forecasting, and other scientific time-evolving systems.