Invited Talks

박경문 교수
(고려대)

Biography
2026-현재	고려대학교 인공지능학과 부교수
2025-2026	고려대학교 인공지능학과 조교수
2021-2025	경희대학교 컴퓨터공학부 조교수
2020-2021	한국전자통신연구원 연구원

Towards Privacy-Preserving Generative Intelligence (45분)

The rapid advancement of generative AI significantly enhances the efficiency and productivity of our daily lives. However, this progress has simultaneously accelerated critical concerns regarding data privacy and the misuse of personal data, such as facial and vocal identities. To address this issue, machine unlearning has emerged as a promising paradigm, enabling models to selectively forget sensitive information without the need for costly retraining from scratch. In this talk, I first introduce the concept of machine unlearning, and then present our group's recent work on machine unlearning-based privacy-preserving techniques for modern generative models. Through this talk, we aim to share insights toward a future of safe generative intelligence where creativity and capability are balanced with robust privacy guarantees.

이동하 교수
(연세대)

Biography
2023-현재	연세대학교 인공지능학과 조교수
2021-2023	University of Illinois at Urbana-Champaign 박사후연구원
2020-2021	포항공과대학교 인공지능연구원 박사후연구원
2020	포항공과대학교 컴퓨터공학(데이터마이닝) 박사
2018-2019	University of Texas Health Science Center at Houston 방문연구원

Towards Personalized AI Agents: From User Understanding to Aligned Behavior (45분)

As AI systems are used by increasingly diverse users, tailoring their responses, decisions, and actions to each individual has become a central challenge for personalized AI. This talk presents our recent efforts toward personalized AI that span the full pipeline. We start from benchmarking personalized behaviors across realistic tasks such as search, shopping, recommendation, and web navigation, then move to modeling user preferences and memory from rich interaction history, and finally to optimizing model behavior with user-specific rewards. Throughout, we discuss how large language models turn these ideas into real-world web and information-retrieval applications, and share empirical studies on how personalized user understanding shapes the evaluation and design of practical AI systems.

최윤석 교수
(Texas A&M University)

Biography
2014-현재	Professor, Texas A&M University
2007-2014	Associate Professor, Texas A&M University
2001-2007	Assistant Professor, Texas A&M University
2017-2019	Corporate Vice President, Samsung Research

Beyond Loss: Ambient Constraints as a Driver of Evolutionary Emergence of Prediction (45분)

Modern machine learning is largely based on the gradient information computed with loss functions, using labels in supervised learning, prediction errors in self-supervised learning, to rewards in reinforcement learning. In this talk, we pay attention to a complementary mechanism for learning: ambient constraints. These include sensory delay, occlusion, energy limits, environmental change, etc. that are prevalent in real-world systems and generally considered as obstacles to be overcome. However, counterintuitively, these obstacles can lead to the emergence of novel functionality such as prediction, without explicitly being optimized for. In this paper, we draw from a line of published works that exhibit this kind of property in evolutionary simulations to argue that ambient constraints can serve as an indirect influence on the learning system to give rise to novel function. This perspective complements explicit gradient-based learning, asking what implicit constraints necessitate the desired capability (such as prediction) and its emergence, rather than asking how to optimize explicitly for the capability.

김태현 교수
(한양대)

Biography
2018-현재	한양대학교 컴퓨터소프트웨어학부 조교수/부교수
2016-2018	막스플랑크연구소 박사후 연구원
2011-2016	서울대학교 공과대학 전기정보공학부 공학박사
2025-현재	IEEE TPAMI Associate Editor
2026-현재	JIPS Editor in Chief
2022-현재	CVPR, ICCV, WACV, ICPR, BMVC, ACCV, NeurIPS, AAAI 등 컴퓨터비전 및 머신러닝 분야 최우수 국제학술대회 Area Chair

Generating Training Datasets under Practical Constraints (45분)

Generative models have demonstrated remarkable success in producing high-quality images and videos. In this presentation, I will introduce our recent research on generating training datasets under practical constraints, such as limited data availability and unpaired datasets. In particular, I will present our latest work on synthesizing realistic low-quality images with authentic degradation artifacts, including image noise and blur, which can be effectively used to train image restoration models for recovering high-quality images.

전광성 교수
(POSTECH)

Biography
2026-현재	포항공과대학교 인공지능대학원/컴퓨터공학과 부교수
2019-2025	University of Arizona, Computer Science 조교수
2018-2019	Boston University, Hariri Institute 박사후연구원
2015-2018	UW-Madison, Institute for Discovery 박사후 연구원
2009-2015	UW-Madison Computer Science 박사
2026-현재	ICML Senior Area Chair

Frontiers of Offline Interactive Machine Learning: From Contextual Bandits to LLM Alignment (45분)

At the heart of modern machine learning lies a fundamental challenge: how can an intelligent system not just learn from data but also decide which data to collect for learning? This is the essence of interactive machine learning (IML) -- a paradigm that encompasses reinforcement learning, contextual bandits, and active learning. Recently, the offline version of IML has gained popularity because the traditional online version often cannot be run due to real-world constraints. In this talk, I will show two recent advances in offline IML problems. First, I will discuss the contextual bandit problem that has applications in recommendation systems. I will show how an improved confidence bound for [0,∞)-valued random variable translates into a superior learning algorithm, both in theory and practice. Second, I will show that the LLM alignment problem is an instance of offline IML and that existing training objectives for it lack theoretical justification, leaving us wondering if they are the right ones to use. As such, I will present a novel theoretical framework for alignment from which three different alignment algorithms are derived along with theoretical guarantees, which is a strong form of justification. Surprisingly, two of them are very similar to existing algorithms called Direct Policy Optimization (DPO) and reinforcement learning from human feedback (RLHF), respectively. Together with our theoretical guarantees, our work can be seen as providing theoretical justifications to DPO and RLHF, with minor corrections. Furthermore, our theory confirms the existing empirical finding that RLHF performs better than DPO. I will conclude with empirical results and exciting future research directions.

서지원 교수
(서울대)

Biography
2025-현재	서울대학교 전기정보공학부/첨단융합학부/협동과정인공지능전공 부교수
2017-2025	한양대학교 컴퓨터소프트웨어학부 부교수
2016-2017	UNIST 전기컴퓨터공학부 조교수
2015	핀터레스트 소프트웨어 (리서치) 엔지니어
2014	LinkedIn 방문연구원
2015	Stanford University EECS 공학박사

Exploiting Speculation in LLM Inference (45분)

Speculative decoding (SD) is an optimization technique for reducing the latency of LLM inference by using a small draft model to propose tokens that are then verified by a larger target model. In this talk, I will present two recent projects from our group that build on SD: the first improves serving throughput by integrating speculation into the inference system, while the second, Speculative Verification, improves efficiency by predicting token acceptance and reducing unnecessary verification cost.

백성용 교수
(한양대)

Biography
2022-현재	한양대학교 데이터사이언스학부/인공지능학과 조교수
2015-2022	서울대학교 공과대학 (전기정보공학/인공지능) 공학박사
2019-2020	Meta Reality Labs (formerly, Facebook Reality Labs) Research Intern

Beyond Parameter Interference: Fine-Tuning, Routing, and Recovery in Model Merging (45분)

The pretraining–finetuning paradigm has produced an explosion of task-specific experts, motivating efforts to merge them into a single multi-task model—yet parameter interference keeps merged models from matching the individual experts. This talk reframes interference as an offset added to each expert's parameters and surveys three works that attack it at three points in the merging pipeline. SAFT uses sharpness-aware fine-tuning to find flat minima, yielding experts that are intrinsically robust to interference. SiM sidesteps interference with training-free task classification based on SVD-subspace projection residuals, routing each input to the right compressed task vector. ReTeX directly predicts and removes the offset, recovering each expert from a single merged checkpoint. Notably, SAFT's perturbation and ReTeX's recovery offset are the same quantity with opposite sign, and SiM and ReTeX share the same projection-residual task identifier—together reaching near-expert performance (95–99%) without router training, task IDs, or storing all experts.

송길태 교수
(부산대)

Biography
2016-현재	부산대학교 정보컴퓨터공학부 교수
2026-현재	부산대학교 정보의생명공학대학장
2020-현재	부산대학교 인공지능융합연구센터장
2025-현재	과학기술정보통신부 AI대학원협의회 이사
2022-현재	한국정보과학회 부회장 (‘24~’25) / 이사 (’22, ‘26)
2022-현재	한국정보과학회 인공지능소사이어티 부회장
2019-현재	부산대학교병원 융합의학기술원 의료인공지능연구실장
2026-현재	한국생명정보학회 이사
2018-현재	ISMB 프로그램 운영위원
2025	RECOMB CG 2025 워크샵 PC chair
2023, 2025	IEEE BigComp PC co-chair

Bridging Structure and Interpretability: From Hypergraph Knowledge Inference to Adaptive Multi-Branch Decision Trees (45분)

As machine learning models are increasingly deployed in high-stakes domains, the dual demands for powerful representation learning and transparent interpretability have become paramount. This talk presents two distinct, novel architectures designed to tackle these challenges across both complex relational data and tabular prediction datasets.

First, we introduce the Hypergraph Interaction Transformer (HIT), a deep representation learning model engineered to infer knowledge across multiple known ontologies. By leveraging hypergraph structures of heterogeneous entities and employing attention-based learning, HIT successfully captures intricate, multi-entity relationships to discover novel knowledge and identify missing associations with state-of-the-art performance and explicit explainability.

Second, shifting the focus to interpretable tabular modeling, we address the under-expressiveness of conventional binary decision trees under strict maximum-depth constraints. We present the Multi-Branch Neural Decision Tree with Adaptive Pruning (MBNDT), a novel framework for shallow-depth tree induction. Trained end-to-end with differentiable multi-way splits, MBNDT learns ordered thresholds over selected features alongside an adaptive branch mask that dynamically optimizes its effective arity. The resulting model is then converted into a deterministic single-path tree for highly efficient, human-readable inference.

Together, these methodologies illustrate how leveraging advanced structural representations and dynamic routing mechanisms can drastically push the frontiers of Explainable AI (XAI) without sacrificing predictive power.