Presentation Details
(DAY 2)
Offline at Auditorium(B1), Samsung Library
10:00~10:50
Toward Model Robustness and Overcoming Data Scarcity for Real-world Visual Intelligence
허재필 교수
Recently, computer vision technology has been advancing at a rapid pace. However, for these technologies to be successfully implemented in the real world, many challenges must be addressed. In this presentation, we will focus on two primary issues: enhancing the robustness of models and overcoming data scarcity. To begin with, we will introduce a self-supervision method designed to complement supervised learning. This approach aids in extracting rich features, thereby facilitating enhanced model robustness. Subsequently, we will present an open-set sample simulation technique, a strategy devised to guide models in effectively handling new class samples that were not encountered during the training phase. Lastly, we delve into strategies to overcome data scarcity, a significant hurdle in this domain. We will introduce a technique that leverages a limited dataset to adapt pretrained GANs. Moreover, we will discuss an unsupervised learning algorithm devised to address the high labeling costs of semantic segmentation.
Physics-Informed Machine Learning
박은병 교수
Physics-informed neural networks (PINNs) have recently emerged as promising data-driven PDE solvers showing encouraging results on various PDEs. However, there is a fundamental limitation of training PINNs to solve multi-dimensional PDEs and approximate highly complex solution functions. The number of training points (collocation points) required on these challenging PDEs grows substantially, but it is severely limited due to the expensive computational costs and heavy memory overhead. To overcome this issue, we propose a network architecture and training algorithm for PINNs. The proposed method, separable PINN (SPINN), operates on a per-axis basis to significantly reduce the number of network propagations in multi-dimensional PDEs unlike point-wise processing in conventional PINNs. We also propose using forward-mode automatic differentiation to reduce the computational cost of computing PDE residuals, enabling a large number of collocation points (>10^7) on a single commodity GPU. The experimental results show drastically reduced computational costs (62x in wall-clock time, 1,394x in FLOPs given the same number of collocation points) in multi-dimensional PDEs while achieving better accuracy. Furthermore, we present that SPINN can solve a chaotic (2+1)-d Navier-Stokes equation significantly faster than the best-performing prior method (9 minutes vs 10 hours in a single GPU), maintaining accuracy. Finally, we showcase that SPINN can accurately obtain the solution of a highly nonlinear and multi-dimensional PDE, a (3+1)-d Navier-Stokes equation.
11:00~12:15 [초청강연]
Information Theory in Computer Vision: NeRF, Debiasing, and Generative Models
한보형 교수 (서울대학교)
Abstract: Information theory is a crucial toolset for quantifying and measuring information, and is widely used in various fields of study including machine learning and its applications. This talk introduces how the basic concepts of information theory can be utilized for problems in computer vision and machine learning, specifically in algorithm 3D view synthesis, debiasing, and learning generative models, based on deep learning. In particular, it deals with how to approximately compute intractable quantities through mathematical analysis and learning.
Biography: Bohyung Han is a Professor in the Department of Electrical and Computer Engineering at Seoul National University, Korea. He received a Ph.D. degree from the Department of Computer Science at the University of Maryland, College Park, MD, USA, in 2005. He served or will be serving as a conference organizing and technical program committee member including a TPC Vice-Chair in ICASSP 2024, a Senior Area Chair in CVPR, NeurIPS, and ICLR, and an Area Chair in CVPR, ICCV, ECCV, NIPS/NeurIPS, ICLR, and IJCAI, a General Chair in ACCV 2022, a Tutorial Chair in ICCV 2019, a workshop chair in CVPR 2021, and a Demo Chair in ECCV 2022. He is also an Associate Editor at TPAMI. He received the Google AI Focused Research Award in 2018, and his research group won the Visual Object Tracking (VOT) Challenge in 2015 and 2016.
13:30~14:45
연구실 소개 및 우수 논문 포스터 행사
(상세내용은 메뉴의 Open-Lab 참조)
https://sites.google.com/view/skkuai2023/presentation-details/openlab-offline-sep-22?authuser=0
패널 토론
(성대오비: 성공적인 대학원생활을 위한 5가지 비결)
패널 : 박호건, 이호준, 차수영, 황성재 교수
좌장 : 이종욱 교수
15:00~16:15
Bridging the Gap: Domain Transfer and Multimodal Learning
홍성은 교수
In this talk, we will focus on two main areas in artificial intelligence: Domain Adaptation and Multimodal Learning. We'll start by looking at how Domain Adaptation helps AI models stay accurate even when the data changes. This technique allows models to adapt without needing new labels. Next, we will discuss Multimodal Learning, which uses multiple types of data, like text and images, to make better predictions. Both these technologies help AI perform well in diverse environments and are crucial for advancing the field.
Trends and Beyond in Data-Driven Reinforcement Learning
김유성 교수
Reinforcement learning is a research field where AI agents learn optimal action policies through trial and error in given environments. Successful examples range from AlphaGo and robot locomotion & manipulation to nuclear fusion plasma control. However, applying reinforcement learning directly to real-world environments can be challenging due to high costs and safety concerns. In this talk, we introduce data-driven reinforcement learning, which has received considerable attention both in academia and industry. Also known as Offline RL, this approach utilizes log data accumulated from humans or rule-based programs to enhance decision-making capabilities through reinforcement learning. One key advantage is that it doesn't require direct interaction with the environment, offering benefits in terms of cost and safety. Nevertheless, Offline RL faces the challenge of data distribution shift, where the data used for learning may differ from the experiences encountered during real evaluation. Addressing this issue is the core of fundamental research in Offline RL. Furthermore, by training on not only narrow & closed task data but also incorporating vast, diverse task data, we can expect to create a generalized learning model that can share knowledge to a broader range of environments.
Privacy-Preserving Decentralized Machine Learning
김형식 교수
In the modern digital era, data is often equated to the value of oil. As machine learning models increasingly depend on vast amounts of data, the traditional centralized model of data processing faces numerous challenges. Real-world applications, especially in the domain of IoT, exhibit a decentralized nature of data. This decentralization poses high data transmission costs and serious privacy concerns. For instance, self-driving cars generate terabytes of data daily, making centralization inefficient. Additionally, increasing public awareness and stringent regulations, like the GDPR, emphasize the importance of data privacy. This talk delves deep into the concept of 'Privacy-Preserving Decentralized Machine Learning.' We will explore advanced techniques like Federated Learning and Split Learning, which promise efficient learning while ensuring user data privacy. Please join us to understand the challenges, potential solutions, and future trajectory of machine learning in a decentralized world.
16:30~17:30 [초청강연]
NC 생성형 AI 기술과 VARCO LLM
박재현 박사 (NC Soft)
초록 : Chat GPT를 시작으로 거대언어모델을 기반으로 한 생성형 AI 기술은 빠르게 우리의 일상 생활에 도입이 되고 있습니다. 생성형 AI 기술이 그 동안 사람과 사람, 사람과 컴퓨터 간의 상호작용 또는 정보 처리의 방식을 근본적으로 바꿀 것이며, 이미 사회의 여러 분야에서 생성형 AI 기술을 활용하여 기술적인 진보를 만들어 내기 위해 다양한 시도들이 이루어지고 있습니다. 이번 발표에서는 NCSoft에서 개발한 거대언어모델과 관련 기술들에 대한 소개하면서, 거대언어모델이 가진 특성을 이해하고 이를 활용하기 위해선 우리가 어떻게 접근해야 하는지 논의해 보고자 합니다.
소개: Jae-Hyun Park is the director at NCSoft. He received a Ph.D. degree from the Department of Computer Science at the University of Massachusetts Amherst, MA, USA in 2014. He received his BS and MS degrees from the Department of Computer Science at Korea University in 2005 and 2007, respectively. He is interested in connecting our billions of users to campaigns relevant to users' interests. By impressing relevant campaigns, we aim to make users not just click the link of a landing page but also do some meaningful actions such as watching the entire video, participating in a survey, and installing the app.