Invited Talks

김선주 교수
(연세대)

Biography

Seon Joo Kim received the BS and MS degrees from Yonsei University, Seoul, Republic of Korea, in 1997 and 2001. He received the Ph.D. degree in Computer Science from the University of North Carolina at Chapel Hill in 2008. He is currently an Underwood Distinguished Professor in the Department of Computer Science, Yonsei University. He has served as a Senior Area Chair for CVPR 2023 and an Area Chair for CVPR 2016, 2018, 2020-2022. He will also serve as an Area Chair for ICCV 2023 and NeurIPS 2023. He is also serving as an Editorial Board member of IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) and International Journal of Computer Vision (IJCV). He is a Senior Member of IEEE. His research interests include computer vision, specifically in computational photography, video understanding, and video processing.

Handling the Discrepancy in training and inference for Video Processing (45분)

Significant progress has been made in image processing and understanding in the last decade, thanks to deep learning. While progress has also been made in video related tasks in computer vision, handling videos, especially long videos, remains challenging. In this talk, I will mainly talk about two video related tasks - video instance segmentation (VIS) and online action detection (OAD). In both tasks, I argue that the biggest bottleneck in current approaches is the discrepancy between training and inference. I will introduce ways to effectively bridge the gap for VIS and OAD, and show that we can reach the state-of-the-art by simply training the videos better.

이기민 교수
(KAIST)

Biography

Kimin Lee is an assistant professor at the Graduate School of AI at KAIST. He is interested in the directions that enable scaling deep reinforcement learning (RL) to diverse and challenging domains — reinforcement learning from human feedback (RLHF), unsupervised reinforcement learning, and representation learning (transformers, self-supervised learning) for RL. Before joining KAIST, Kimin Lee was a research scientist at Google Research in Mountain View. He completed his postdoctoral training at UC Berkeley (advised by Prof. Pieter Abbeel) and received his Ph.D. from KAIST (advised by Prof. Jinwoo Shin). During his Ph.D., he also interned and collaborated closely with Honglak Lee at the University of Michigan.

Reinforcement Learning from Human Feedback for Aligning Text-to-Image Models (45분)

Learning from human feedback has been shown to improve text-to-image models. These techniques first learn a reward function that captures what humans care about in the task and then improve the models based on the learned reward function. Even though relatively simple approaches (e.g., rejection sampling based on reward scores) have been investigated, fine-tuning text-to-image models with the reward function remains challenging. In this work, we propose using online reinforcement learning (RL) to fine-tune text-to-image models. We focus on diffusion models, defining the fine-tuning task as an RL problem, and updating the pre-trained text-to-image diffusion models using policy gradient to maximize the feedback-trained reward. Our approach, coined DPOK, integrates policy optimization with KL regularization. We conduct an analysis of KL regularization for both RL fine-tuning and supervised fine-tuning. In our experiments, we show that DPOK is generally superior to supervised fine-tuning with respect to both image-text alignment and image quality.

원정담 교수
(서울대)

Biography

Jungdam Won is an assistant professor in the Department of Computer Science and Engineering at Seoul National University in Korea. Before joining SNU, he worked for 3.5 years as a research scientist at Meta (formerly Facebook) AI in the Creative AI team led by Prof. Jessica Hodgins at Carnegie Mellon University. He earned his Ph.D. and B.S. from Seoul National University in 2017 and 2011, respectively. His current research interests include designing controllers for virtual/physical agents, understanding interactions between multiple agents, and developing user-friendly methods that bridge the gap between users and their virtual personas, utilizing various machine learning approaches, motion capture, and optimization techniques. In terms of academic services, he has served as a committee member for prestigious international conferences, including ACM SIGGRAPH/SIGGRAPH Asia (technical paper program and course program), Eurographics, Pacific Graphics, Motion Interaction and Games, CASA, and IEEE AIVR.

Human Motion Generation via Physics Simulation and Its Applications (45분)

Generating 3D human motion is a key research topic in computer graphics, vision, robotics, and medicine, with applications in film-making, computer games, the Metaverse, human-robot collaboration, and gait analysis for diagnosis. Despite recent AI advances, data-driven machine learning approaches still exhibit many artifacts. In this talk, I'll introduce a physics simulation-based paradigm for 3D human motion generation. The talk will start by discussing its potential to overcome existing challenges in 3D human motion generation. It will then delve into several methods of generating human motions based on physics simulation from various sources, including videos, IMUs, 3D markers, text, and more, following the order of recent advances.

전상률 교수
(부산대)

Biography

Sangryul Jeon received the B.S., and Ph.D. degrees from the School of Electrical and Electronic Engineering, Yonsei University, Seoul, South Korea, in 2016 and 2022, respectively. From 2022 to 2023, he was a Postdoctoral Researcher with University of California, Berkeley, USA and University of Michigan, USA. Since 2023, he has been with the Department of Computer Science and Engineering, Pusan National University, South korea. His current research interests include visual correspondence, unsupervised representation learning and object discovery.

Unsupervised Object Discovery (45분)

We consider learning to discover and segment objects from a collection of unlabeled images. Our key insight is that object wholes can be discovered effectively across similar images by analyzing their structural grouping: Not only similar foreground parts in two images attract, but varied foreground parts in one image also commonly repel backgrounds in the other image, thereby binding heterogeneous parts of an object into a coherent whole. We capture the nexus of attraction and repulsion within and across similar images in a weighted graph where full objectness is revealed by maximizing intra-image attraction and inter-image repulsion. We further distill these results in a single-image feature learning model to improve the feature towards whole object discovery. Our method delivers state-of-the-art performance on 9 common benchmarks for unsupervised object discovery, saliency detection, and video object segmentation.

강재우 교수
(고려대)

Biography
2021-현재	㈜아이젠사이언스 CEO/Founder
2006-현재	고려대학교 컴퓨터학과 교수
2003-2006	North Carolina State University 컴퓨터학과 조교수
2003	University of Wisconsin-Madison 전산학 박사
2000-2001	WISEngine Inc., CTO/Founder
1997-1998	Savera Systems Inc., 연구원
1996-1997	AT&T Labs Research, 연구원
1996	University of Colorado at Boulder 전산학 석사
1994	고려대학교 전산학 학사

Scientific Hypothesis Generation and Large Language Models: From the Perspective of Drug Target Discovery (45분)

Navigating the intricate landscape of drug development requires more than a decade of intense research, an investment of roughly $2.6 billion per drug, and regulatory hurdles. The escalating costs threaten the pharmaceutical industry's sustainability, raising concerns about diminishing returns in the near future.

Amidst these challenges, the potential of artificial intelligence (AI) as a game-changer has garnered attention. Can AI truly be the answer to surging costs and inefficiencies? While AI has displayed promising strides in various facets of drug development, bridging the gap between its theoretical promise and real-world application remains a work in progress.

In this session, we will discuss the contributions AI can make throughout the drug development process and the current challenges AI technologies need to address. A particular focus will be on target discovery, the critical initial step in drug development. This phase involves formulating hypotheses about the disease mechanism and identifying potential targets that can inhibit this mechanism. We will examine the role and capabilities of large language models in this context, drawing on the speaker's recent research.

김태현 교수
(한양대)

Biography

Tae Hyun Kim is an assistant professor in the Department of Computer Science at Hanyang University, where he has led the Visual Intelligence Laboratory since 2018. With a focus on solving fundamental and challenging low-level computer vision problems such as super-resolution, deblurring, and denoising, his research aims to advance the field. Tae Hyun Kim received his BS and MS degrees from KAIST and his Ph.D. degree from Seoul National University under the supervision of Prof. Kyoung Mu Lee, specializing in computer vision. After completing his Ph.D., he was a postdoctoral fellow at the Max Planck Institute in Germany, where he worked with Prof. Bernhard Schölkopf and Dr. Michael Hirsch for two years.

Flow based generative models (45분)

Recently, generative models have received considerable attention in various research domains, ranging from applied fields such as natural language processing and computer vision to pure theoretical disciplines such as chemistry and physics. While numerous approaches based on generative adversarial networks and diffusion models have been extensively explored, normalization flow has received comparatively less attention, and its potential for addressing the generation problem remains underappreciated. In this talk, I will demonstrate the remarkable capabilities of normalizing flows in the areas of generation and density estimation, and present our recent research results. By highlighting the promising results achieved, I will emphasize the potential of normalizing flows in the field of computer vision.

윤철희 교수
(KAIST)

Biography

Chulhee “Charlie” Yun is an assistant professor at KAIST Kim Jaechul Graduate School of AI, where he directs the Optimization & Machine Learning Laboratory. He finished his PhD from the Laboratory for Information and Decision Systems and the Department of Electrical Engineering & Computer Science at MIT, under the joint supervision of Prof. Suvrit Sra and Prof. Ali Jadbabaie. Charlie’s research spans optimization and machine learning theory, with the driving goal of bridging the gap between theory and practice.

Shuffling-based Stochastic Optimization Methods: Bridging the Theory-practice Gap (45분)

Stochastic finite-sum optimization problems are ubiquitous in many areas such as machine learning, and stochastic optimization algorithms to solve these problems are actively studied in the literature. However, there is a major gap between practice and theory: practical algorithms shuffle and iterate through component indices in a sequential manner, while most theoretical analyses of these algorithms assume uniformly sampling the indices in an i.i.d. manner. In this talk, I will talk about recent research efforts to close this theory-practice gap, focusing on the progress in the convergence analysis of shuffling-based optimization methods. We will first consider stochastic gradient descent (SGD) with random reshuffling; we will then briefly talk about some recent progress on sampling schemes faster than random reshuffling.

전병곤 교수
(서울대)

Biography

Byung-Gon Chun is the CEO and founder of FriendliAI, a global leader of generative AI serving engine technology, and Professor in the Computer Science and Engineering Department at Seoul National University (SNU). Chun is a world-renowned scholar in the areas of systems for machine learning. During his tenure at SNU, Chun was a Visiting Research Scientist at Facebook. Before joining SNU, Chun was a Principal Scientist at Microsoft. Prior to that, Chun worked at Yahoo! Research, Intel Research, and International Computer Science Institute. Chun holds a Ph.D. in Computer Science from the University of California, Berkeley, an M.S. in Computer Science from Stanford University, and a B.S. and an M.S. from the Electronic Engineering Department at SNU. He has published numerous papers in the premier systems, machine learning, and database conferences such as OSDI, NSDI, EuroSys, ATC, NeurIPS, ICML, VLDB, and SIGMOD. He is the recipient of the 2021 EuroSys Test of Time Award and the 2020 ACM SIGOPS Hall of Fame Award, which recognize the most influential systems papers that were published at least ten years in the past. He also received research awards from Google, Amazon, Facebook, and Microsoft.

Supercharging large language model (LLM) serving with FriendliAI's PeriFlow (45분)

Large language models (LLMs) trained for generation tasks have recently attracted huge interest, emphasizing the need for serving such models very efficiently. However, existing solutions for inference serving underperform on these workloads since they were not designed with them in mind. In this talk, I will share both my research and industry experience of serving LLMs at FriendliAI. In particular, I will present various LLM workloads and optimizations to tackle the challenges of serving LLMs with low latency and high throughput. PeriFlow, our gen AI serving engine, which was born out of our Orca research published in OSDI '22, significantly outperforms other solutions. PeriFlow is available for use at friendli.ai.

김선 교수
(서울대)

Biography
2022-현재	목암생명과학연구소 소장
2011-현재	서울대학교 컴퓨터공학부 교수
2011–2021	서울대학교 생물정보연구소 소장
2009–2011	미국 인디애나대학교 School of Informatics and Computing 학과장
2001–2011	미국 인디애나대학교 School of Informatics and Computing 조교수, 부교수
1998–2001	미국 듀퐁중앙연구소 선임 연구원
1997	아이오와대학교 컴퓨터과학 박사
1987	한국과학기술원 컴퓨터과학 석사
1985	서울대학교 계산통계학 학사

Exploring chemical, genetic and disease space with AI technologies (45분)

In this talk, I will discuss about how AI and deep learning technologies can be used for predicting patient mortality, drug response, toxicity, and repositioning. Before getting into details, I will briefly talk about essential concepts, such as embedding or representations, distance and distributions, that are frequently used for constructing AI models. Then, in part 2 of my talk, I will discuss our recent work on analyzing EHR data for predicting patient mortality and another work on how drug repositioning can be inferred from knowledge graphs that consist of drug, gene, and disease. Part 3 of my talk will focus on the gene level modeling, specifically how AI technologies can be used for modeling drug response, toxicity, and disease subtypes using transcriptome and multi-omics data.

안성수 교수
(POSTECH)

Biography

Sungsoo Ahn is an assistant professor in the Department of Computer Science and Engineering and Graduate School of Artificial Intelligence and a member of the Machine Learning Lab at POSTECH. Prior to this, he was a postdoctoral research associate under the supervision of Prof. Le Song and Prof. Eric Xing at MBZUAI. He received his Ph.D. under Prof. Jinwoo Shin at KAIST. He is broadly interested in machine learning for combinatorial, graph-structured, or geometric data with applications to drug design, material design, and operational research.

Opportunities and Challenges in Deep Graph Generative Models (45분)

Graphs are abstract representations of the relation between objects that find applications in representing structures such as molecules, circuits, and social networks. Hence, learning a graph generative model is a fundamental problem across various domains, including social network analysis, drug and material design, and privacy-aware training of graph neural networks. In this talk, I will discuss the potential and the main challenges of using deep neural networks for graph generation. I will also introduce some of the notable recent works such as the development of diffusion-based models or new graph representations for graph generation.

김주연 교수
(UNIST)

Biography

Jooyeon Kim is an assistant professor at UNIST. His research in machine learning (ML) focuses on building and promoting interactive systems within which humans and machines (AI agents) collaborate, cooperate, and coordinate through verbal and non-verbal communicative signals. The multi-modal nature of such interactive systems leads his research trajectory to revolve around the research areas including natural language processing (NLP), data mining, human-computer interaction (HCI), and optimization. He earned his Ph.D. and M.S. from the Korea Advanced Institute of Science and Technology (KAIST) and his B.E. from the University of Tokyo (東京大学). Since his graduate studies, he has been involved in a startup thingsflow, where he developed chatbot systems, virtual humans, image recognition, and large language models (LLMs). In 2020-2021, he was a researcher at Microsoft Research Cambridge (MSRC). In 2021 — 2023, he was a postdoctoral researcher at RIKEN (이화학연구소: 理化学研究所). During graduate studies, he interned at the Max Planck Institute for Software Systems (MPI-SWS) and Microsoft Research Cambridge (MSRC).

Cooperative Policy Learning in Transformer World Models (45분)

How can we infuse an artificial agent with cooperative behavior? Previous approaches have adopted centralized training, i.e., putting multi-agents together in the same environment throughout multiple episodes, or imitation learning, i.e., resorting to human exhibitions. In this work, we demonstrate how transformer world models facilitate the emergence of a cooperative policy of an agent with minimal real-environment interactions and no human supervision. An agent first learns to envision a world model from the random movements of oneself and the partner agents in the real environment. Inside the world model, the agent coordinates its actions in response to the actions made by the others. The agent can effectively derive a cooperative policy because it can control the actions of itself as well as that of the other imaginary agents. During the process, policy learning is done solely inside the world model. Real-environment interactions are required only to train the world model with respect to new experiences. For the experiments, we design a 3D environment with egocentric pixel-based image observations that may or may not contain partner agents. We evidence the policy learned inside the world model being executed seamlessly within the real environment, allowing the agent to cooperate with humans.