패턴인식/기계학습 세션

정종헌 교수
(고려대)

Biography
2024-현재	고려대학교 정보대학 인공지능학과 조교수
2023-2024	KAIST 정보전자연구소 박사후연구원
2021-2022	AWS AI Applied Scientist Intern
2016-2017	XBrain Inc. Machine Learning Engineer
2017-2023	KAIST 전기전자공학 공학박사
2012-2017	KAIST 수학/전산학 이학학사

Safety Issues in Generative AI at Scale (90분)

Recent breakthroughs in generative AI have been largely driven by scaling laws, unlocking unprecedented capabilities across vision, language, and multimodal domains. However, scaling is inherently difficult to control and cannot be assumed to bring only benefits. This talk will explore several key safety risks emerging from large-scale generative AIs, including issues of fairness, robustness, and privacy. It will also highlight recent research efforts across these areas, covering both technical mitigation strategies and open challenges that remain for building trustworthy generative systems.

김건희 교수
(서울대)

Biography
2015-현재	서울대학교 컴퓨터공학부 교수
2018-현재	(주)리플에이아이 대표
2013-2015	Disney Research, 박사후 연구원
2009-2013	Carnegie Mellon University 컴퓨터과학 박사

Safe and Trustworthy Agents (90분)

Recently, large-scale language models have been driving a revolution in artificial intelligence and rapidly evolving into the agents capable of performing a wide range of intelligent tasks. This lecture explores the fundamentals of LLM-based agent technologies and introduces some of the most noteworthy recent models. In particular, we will delve into safety and trustworthiness issues of these emerging agents that are being developed to carry out diverse digital tasks through GUIs and to function as physical AI, much like humans.

주재걸 교수
(KAIST)

Biography
2020-현재	KAIST 김재철AI대학원 부교수/교수, 석좌교수
2015-2020	고려대학교 컴퓨터학과 조교수/인공지능학과 부교수
2013-2019	UNIST 전기전자컴퓨터, 조교수/부교수
2011-2015	Georgia Tech, Research Scientist
2005-2013	Georgia Tech, 전기공학부 석사, 계산과학및공학부 박사
1997-2001	서울대학교 전기공학부 학사

Sparse Autoencoders for Interpretable Vision Encoders in VLMs (90분)

Sparse autoencoders (SAE) have recently shown its potentials in interpreting the inner-workings of large language models. However, its applications in computer vision domains and multi-modal models have been under-explored. In this talk, I will introduce some of my recent research on this direction, as well as its interesting findings. In detail, this talk will give a comprehensive overview on how the sparse autoencoders work so that we can interpret the large language models. Next, I will introduce how we can apply the SAE in vision encoder models and in particular, what they learn during finetuning for a given target task. Also, I will discuss the extension of SAE in the context of model debiasing. Finally, the talk will conclude with future research directions.

이화란 교수
(서강대)

Biography
2025-현재	서강대학교 조교수
2021-2025	NAVER AI Lab, Lead Research Scientist
2018-2021	SK T-Brain, Research Scientist
2013-2018	KAIST 전기 및 전자공학과 박사
2008-2012	KAIST 수리과학과 학사

Trustworthy LLMs (90분)

In this tutorial, I introduce recent evaluation benchmarks with socio-cultural awareness for large language models. In terms of safety including social bias and values, I briefly present the benchmark construction framework of SQuARe, KoSBi, and KoBBQ. Then, I discuss the vulnerability of multilingual LLMs and red-teaming attack method, namely code-switching red-teaming (CSRT). Finally, I introduce the efficient multilingual transfer method using code-switching curriculum learning.

최재식 교수
(KAIST)

Biography
2019-현재	KAIST 김재철AI대학원 부교수/교수, 석좌교수
2019-현재	인이지, 대표이사
2013-2019	UNIST 전기전자컴퓨터, 조교수/부교수
2013-2013	로렌스 버클리 연구소 박사후과정
2005-2012	일리노이 대학교 어바나 샴페인 컴퓨터과학 박사
1997-2004	서울대학교 컴퓨터공학 학사

Investigating the Internal Mechanisms of Deep Neural Networks (90분)

As complex artificial intelligence (AI) systems such as deep neural networks are used for many mission critical tasks such as military, finance, human resources and autonomous driving, it is important to ensure the safe use of such complex AI systems. In this talk, I will present recent advances to clarify the internal decision of deep neural networks. Moreover, we will overview approaches to automatically correct internal nodes which incur artifacts or less reliable outputs. Furthermore, we will investigate the reasons why some deep neural networks and large language models include not-so-stable internal nodes.

이성윤 교수
(한양대)

Biography
2023-현재	한양대학교 컴퓨터소프트웨어학부 조교수
2021-2023	고등과학원 AI기초과학센터 Research Fellow
2021	서울대학교 수리과학 이학박사

To drive AI as we desire (90분)

2012년 AlexNet으로 대표되는 딥러닝의 성공 직후, 2013년 adversarial example이라는 문제점이 등장했다 (ICLR 2014에서 발표되었지만, 2012년말 소수에게는 이미 알려졌다). 이후 10년간 많은 연구가 진행되었고, 이러한 역할을 인정받아 adversarial example 연구는 2024년 ICLR Test of Time Award도 받았다. 하지만 현재는 어떠할까? 해당 방향 연구가 여전히 의미있는 방향일까? 우리는 해당 방향 이외의 LLM 또는 diffusion model의 다른 safety issues, 그리고 feature learning 특성에 대해서 살펴본다.