Loading...

학술행사

Korean AI Association

학술행사

  >   학술행사   >   국내학술대회

국내학술대회

연사 및 초록 (16일)
 
황원석 교수(서울시립대학교)
 
Title: Tutorial on Large Language Models
 
Abs:

Large language models (LLMs) have demonstrated remarkable performance across diverse domains. For instance, in the legal domain, GPT-4 successfully passes the uniform bar exam. However, the same model fails to pass the Chinese lawyer qualification test and shows suboptimal performance in various other legal AI tasks. Why LLMs display such divergent performance?

This tutorial lecture delves into the underlying technology of LLMs. Starting with a concise introduction to natural language processing, we will briefly review the basics of Transformer neural architecture and study what happens when the model and data scale. Afterward, we will examine the algorithms that empower LLMs to understand and follow instructions. Finally, we will wrap-up the lecture by covering the recent trends in LLMs.

Bio: 

(현) 서울시립대 인공지능학과 조교수 (2023.8-)
(현) 엘박스 Research Scientist (2021.7-2023.7, 2023.10-)
(전) 네이버 클로바 Research Engineer (2018.3-2021.7)
(전) 고등과학원 Research Fellow (2014.9 - 2018.3)
서울대 물리천문학부 물리학 박사(2010.3-2014.8)
서울대 물리천문학부 물리학 석사(2007.9-2010.2)
서울대 물리학부 학사 (2003.3 - 2007.8)
 
 
권세중 리더(네이버클라우드)
 
Title: Unlocking Efficiency in LLMs: The Role of Model Compression​
 
Abs:
본 강좌에서는 그간 다양하게 연구되어 온 Model Compression 기술들에 대해 소개하고, 해당 기술들이 Large Langauge Model 기반 AI Service들의 운용비용을 낮추는 관점에서 어떤 의미가 있는지를 다룬다.
 
Bio: 
권세중은 2009년 KAIST 전산학과를 졸업하고, KAIST 전기및전자공학과에서 Modeling and Simulation을 주제로 박사 학위를 받았다. 삼성리서치에서 3년간 On-device AI를 주제로 모델 압축/최적화 알고리즘을 연구했으며, 지금은 네이버 클라우드 HyperScale AI 조직에서 Model Optimization 팀을 리딩하여, HyperCLOVA-X의 경량화/최적화 업무를 진행해오고 있다. CVPR/ICML/ICLR/Neurips/EMNLP 등 여러 딥러닝 학회 뿐만 아니라 DAC/MLSys/SC 등의 시스템/하드웨어 학회에도 다양한 논문을 채택시켰으며, 최근에는 삼성전자-네이버 AI반도체 개발협업 업무를 맡아, 새로운 AI 반도체를 통해 저비용/고효율의 AI 서비스 서빙 시스템을 실현하고자 노력하고 있다.
 
 
 
박찬영 교수(KAIST)
 
Title: Learning on Graph and its Application to Biomedical Data
 
Abs:
In recent times, in the field of chemistry and materials science, Graph Neural Networks (GNNs) have been extensively utilized for modeling the properties and behavior of materials and molecules. GNNs are effective in learning from data with graph structures, making them widely applicable to molecular and crystal structures, which can be easily represented as graphs. GNNs that incorporate both spatial and non-spatial information have shown promising results in predicting properties such as band gaps, formation energies, and reaction energies. In this lecture, I will introduce the fundamental concepts of GNNs and explore examples of their application in predicting chemical reactions.
 
Bio: 
2020 - Present :  Assistant Professor, KAIST
2019 - 2020 : Postdoctoral Research Fellow, University of Illinois at Urbana-Champaign
2019 : Ph.D, POSTECH 
 
 
옥정슬 교수(POSTECH) 
 
Title: Reinforcement learning: basics and applications 
 
Abs:
Reinforcement Learning (RL) has been emerged as a core paradigm for a wide spectrum of applications. This is perhaps because RL framework can be applied for optimizing various objectives, regardless of their differentiability, including a simple scalar and even human preference (e.g., fine-tuning via RL for chatGPT). This talk will provide the fundamentals of RL, elucidating its core concepts and algorithms. Beginning with an overview of the basic principles underlying RL, including Markov Decision Processes (MDPs) and the exploration-exploitation dilemma, key algorithms such as Q-learning, Deep Q-Networks (DQN), and Policy Gradient methods will be introduced. Then, this talk will be concluded with several real-world applications such as game playing, complex visual recognition system, and language models.
 
Bio: 
Jungseul Ok is an associate professor in the Department of Computer Science and Engineering and Graduate School of Artificial Intelligence, and a member of Machine Learning Lab at POSTECH. He completed Ph.D program in  School of Electrical Engineering at Korea Advanced Institute of Science and Technology (KAIST), South Korea, under the supervision of Prof. Yung Yi and Prof. Jinwoo Shin. After graduation, he worked with Prof. Alexandre Proutiere and Prof. Sewoong Oh as a postdoctoral researcher, respectively, in School of Electrical Engineering at KTH, Stockholm, Sweden, and Paul G. Allen School of Computer Science & Engineering, University of Washington, WA, US.