Invited Talks

최재식교수
(KAIST)

Biography
2019-현재	Associate Professor, KAIST
2017-현재	Director, Explainable Artificial Intelligence Center
2013-2019	Assistant/Associate Professor, UNIST
2012	PhD, CS, University of Illinois Urbana-Champaign
2004	BS, CSE, Seoul National University

설명가능 인공지능 연구동향: 딥러닝 내부 분석 및 수정 기술 중심으로 (45분)

Explainable and interpretable machine learning models and algorithms are important topics which have received growing attention from research, application and administration. Many advanced Deep Neural Networks (DNNs) are often perceived as black-boxes. Researchers would like to be able to interpret what the DNN has learned in order to identify biases and failure models and improve models. In this tutorial, we will provide a comprehensive overview on methods to analyze deep neural networks and an insight how those XAI methods help us understand time series data. In addition, we will present a new method to automatically correct the incorrectly trained units in deep neural networks.

서민준교수
(KAIST)

Biography

Minjoon Seo is an Assistant Professor at KAIST Graduate School of AI. He finished his Ph.D. at the University of Washington, advised by Hannaneh Hajishirzi and Ali Farhadi. His research interest is in natural language processing and machine learning, and in particular, how knowledge data can be encoded (e.g. external memory and language model), accessed (e.g. question answering and dialog), and produced (e.g. scientific reasoning). His study was supported by Facebook Fellowship and AI2 Key Scientific Challenges Award. He previously co-organized MRQA 2018, MRQA 2019 and RepL4NLP 2020.

Large Language Models (45분)

In this talk, I will give an overview on recent trend in large language models. I will first begin with classification models such as GPT-1 and BERT and then discuss sequence-to-sequence models such as BART and T5. Then I will move to general-purpose generation models such as GPT-2 and GPT-3. I will also delve into language models with other modalities than language, such as Dall-E (image) and GSLM (speech). I will conclude with their implications and impact in not only research community but also the society as a whole.

홍승훈교수
(KAIST)

Biography

He is an assistant professor at the School of Computing, KAIST. Before joining KAIST, he was a visiting faculty researcher at Google Brain, and a postdoctoral fellow at University of Michigan collaborated with Professor Honglak Lee on topics related to deep learning and its application to computer vision. He received his Ph.D. degree at POSTECH, Korea under the supervision of Professor Bohyung Han. His research interests include machine learning and computer vision. Particularly, he is interested in scaling up machine learning algorithms for visual perception by minimizing human supervision for training. He is also interested in making such algorithms interpretable to humans, allowing users to more easily understand and get involved in the decision making process in ML systems.

A Unified View and Neural Networks for Sets, Graphs, and Hypergraphs (45분)

Graphs are one of the most general representations of the data, covering from the sequence (e.g., language) and grid (e.g., images) to chemical structures (e.g., drugs) and even arbitrary networks (e.g., social network). The importance of processing and understanding graphs has become an important topic in the machine learning community. Yet, models for handling graph-structured data have been developed separately for sets, graphs, and hypergraphs, while attempts to unify these models are in an early stage. In this talk, I will introduce our recent work towards building a unified framework for sets, graphs, and hypergraphs. I will first discuss a general representation of these graph-structured data and our model for processing them, which turned out to be a generalization of the Transformer.

주재걸교수
(KAIST)

Biography

Jaegul is currently an associate professor in the Graduate School of Artificial Intelligence at KAIST. He has been an assistant professor in the Dept. of Computer Science and Engineering at Korea University from 2015 to 2019 and then an associate professor in the Dept. of Artificial Intelligence at Korea University in 2019. He received M.S in the School of Electrical and Computer Engineering at Georgia Tech in 2009 and Ph.D in the School of Computational Science and Engineering at Georgia Tech in 2013, advised by Prof. Haesun Park. From 2011 to 2014, he has been a research scientist at Georgia Tech. During the summer in 2009 and 2010, he worked at National Visualization and Analytics Center (NVAC) in Pacific Northwest National Laboratory. He earned his B.S in the Dept. of Electrical and Computer Engineering at Seoul National University.

Domain Generalization and Out-of-Distribution Detection in Urban-Scene Semantic Segmentation (45분)

Despite the remarkable success of urban-scene semantic segmentation models, they cannot properly handle (1) the domain shift between training and test data nor (2) out-of-distribution data that do not belong to the predefined classes. In this talk, I will introduce our recent studies that addresses the task of domain generalization and out-of-distribution (OOD) detection in urban-scene semantic segmentation. For the former, I will present my recent work called RobustNet [CVPR’21 Oral], which disentangles the domain-specific style and domain-invariant content and removes the style information of the training domain. For the latter, I will present a simple yet effective approach [ICCV’21 Oral] that standardizes the prediction scores of each class significantly different from each other and uses this standardized max logits as a measure for OOD detection. Both approaches achieve the new state-of-the-art in their respective tasks while involving a minimal amount of additional computation cost. I will conclude the talk with future research directions.

김선주교수
(연세대학교)

Biography

Dr. Seon Joo Kim is currently an Associate Professor of Computer Science at Yonsei University. He received B.S. and M.S. degrees in EE at Yonsei University, and Ph.D. in Computer Science at University of North Carolina at Chapel Hill advised by Dr. Marc Pollefeys. He worked with Dr. Michael Brown at National University of Singapore (2009-2011), and as an Assistant Professor at SUNY Korea in 2012. He joined Yonsei University in 2013 and worked as a Visiting Scientist at Facebook (2019-2020). His research includes low level computer vision including color processing and image/video restoration as well as high level video understanding and processing. Dr. Kim has served as an Area Chair for ICCV 2017, CVPR 2016,2018,2020,2021, 2022, and also as an local arrangement chair for ICCV 2019 in Seoul.

Towards practical computer vision systems in super-resolution and video understanding (90분)

While we have reached a new era in computer vision with deep learning, we still have not seen many practical computer vision systems deployed in the real world. In this talk, I would like to present ways to improve practicality of computer vision algorithms. I will specifically focus on the problem of super-resolution and video understanding. In super-resolution, topics include loss function learning, tackling the ill-posedness of SR through adaptive target generation, and fast SR with look-up tables. For video understanding, I will introduce a new task called online temporal action localization, and how to solve the problem through imitation learning.

진소영박사
(MIT)

Biography

She is a postdoctoral associate at the Computer Science and Artificial Intelligence Laboratory (CSAIL) at the Massachusetts Institute of Technology (MIT), working with Aude Oliva. Previously she earned her PhD in College of Information and Computer Sciences (CICS), University of Massachusetts, Amherst (UMass Amherst), where she researched on improving face clustering in videos under Erik Learned-Miller in the Computer Vision Lab. Her research interests are in Computer Vision and Machine Learning.

Cross-Modal Learning (45분)

컴퓨터가 어떻게 하면 동영상을 잘 이해할 수 있을까를 연구하는 Video Understanding에 대해 전반적으로 소개합니다. Cross-Modal Learning은 두 개 이상의 서로 다른 센서로 얻은 정보 (예: 동영상과 그 동영상을 설명한 캡션) 를 이용해서 컴퓨터 모델을 학습시키는 방법을 말합니다. 이러한 Cross-Modal Learning을 Video Understanding 에 어떻게 사용할 수 있는지에 대해, 최근 발표한 두 논문과 함께 설명하겠습니다.

서유진박사
(Brown University)

Biography

Eugene Seo is a postdoctoral research associate in Applied Mathematics at Brown University. She received her Ph.D. degree in Computer Science from Oregon State University, M.S. degree in Computer Science from Korea Advanced Institute of Science and Technology, and B.Eng. degree in Computer Science and Electronic Engineering from Handong Global University. Her research interests include machine learning, data mining, and computational sustainability.

Modeling Species Interactions and Distributions Under Imperfect Detection (45분)

Ecological domains seeking to understand the environment and the behavior of species have received little attention in machine learning (ML), despite the fact that environmental changes have a significant impact on humans as well as ecosystems. In this talk, I will present how ML techniques can be applied to ecological problems. In the first part of the talk, I will show problem formulation for plant-pollinator interaction analysis and the use of latent factor models to understand the underlying mechanisms that determine species interactions. The second part of the talk addresses imperfect detection, which is a pervasive challenge of ecological data. I will present a new link prediction framework customized for ecological networks by combining the Poisson N-mixture model, a latent variable model widely used in statistical ecology for modeling imperfect detection of a single species, with a probabilistic nonnegative matrix factorization model for modeling multiple species interactions. In the last part, I will introduce a computational framework called StatEcoNet, which integrates neural networks with a graphical generative model, to handle model complexity as well as the imperfect detection problem in species distribution modeling.

이민경교수
(University of Texas at Austin)

Biography

Min Kyung Lee is an assistant professor in the School of Information at the University of Texas at Austin. She is affiliated with UT Austin Machine Learning Lab—one of the first NSF funded national AI research institutes, Good Systems—a UT Austin 8-year Grand Challenge to design responsible AI technologies, and Texas Robotics. Previously, she was a research scientist in the Machine Learning Department at Carnegie Mellon University. Dr. Lee has conducted some of the first studies that empirically examine the social implications of algorithms’ emerging roles in management and governance in society. She has extensive expertise in developing theories, methods and tools for human-centered AI and deploying them in practice through collaboration with real-world stakeholders and organizations. She developed a participatory framework that empowers community members to design matching algorithms that govern their own communities. Her current research is inspired by and complements her previous work on social robots for long-term interaction, seamless human-robot handovers, and telepresence robots. Dr. Lee is a Siebel Scholar and has received the Allen Newell Award for Research Excellence, research grants from NSF and Uptake, and five best paper awards and honorable mentions and two demo/video awards in venues such as CHI, CSCW, DIS, HRI and MobiSys. She is an associate editor of ACM Transactions on Human-Robot Interaction. Her work has been featured in media outlets such as the New York Times, New Scientist, Washington Post, MIT Technology Review and CBS. She received a PhD and a MS in Human-Computer Interaction and an MDes in Interaction Design from Carnegie Mellon University and a BS from KAIST.

Responsible and Fair AI: Human-Centered Approaches (90분)

TBD

송길태교수
(부산대학교)

Biography

Giltae Song, Ph.D., is an associate professor in the School of Computer Science and Engineering at Pusan National University (PNU). Before joining PNU, he was a post-doctoral scholar in Prof. J.Michael Cherry’s group at Stanford University. Dr. Song earned a Ph.D. in computer science and engineering at Pennsylvania State University (advised by Prof. Webb Miller), and a bachelor’s degree and a master’s degree in computer science and engineering at Seoul National University. His research focuses on machine learning and data mining specialized for analyzing various biomedical data (e.g. genome sequence data, experimental data for drug discovery, and clinical data in hospitals).

Recommendation systems in biomedicine (45분)

Recommendation system techniques have been applied in various areas such as movie and music recommendations, and product suggestions in on-line stores. Deep neural network architecture has been integrated with traditional matrix factorization. This lecture covers some recent advanced recommender systems based on deep learning as well as common recommendation system techniques such as matrix factorization and latent factor models. In addition, we introduce several types of problems that have been resolved using recommender systems in healthcare and biomedical research.

이슬교수
(아주대학교)

Biography

Sael Lee is currently an Associate Professor in the Dept. of Software and Computer Engineering and the Dept. of Artificial Intelligence. From 2012-2018 she was an Assistant Professor in the Dept. of Computer Science at Stony Brook University Korea. She received her Ph.D. in Computer Science from Purdue University in 2010 and her B.S. in Computer Science from Korea University in 2005. She has published in major journals and proceedings using her publication name, "Lee Sael," in the areas of machine learning for healthcare, tensor analysis/data mining, and bioinformatics.

Deep Trees Models (45분)

How can we improve the performance of classifications of structured data, including the tabular data? Deep neural network models are being widely applied to improve the performance of machine learning problems applied to large-scale unstructured data. However, performance improvements of these models have been limited when applied to the traditional tabular data that still make up for a large percentage of available data. In this talk, I will go over recent models that target tabular data, focusing on tree-based models, which have additional benefits of easiness in interpretability.

류경석교수
(서울대학교)

Biography

Ernest Ryu is an assistant professor in the Department of Mathematical Sciences at Seoul National University. He is an affiliated faculty of the Graduate School of Artificial Intelligence and the Graduate School of Data Science. His current research focus is on optimization and deep learning theory. Professor Ryu received a BS degree in Physics and Electrical engineering with honor at the California Institute of Technology in 2010 and an MS in Statistics and PhD in Computational and Mathematical Engineering at Stanford University in 2016. In 2016, he joined the Department of Mathematics at the University of California, Los Angeles as an Assistant Adjunct Professor. In 2020 he joined the Department of Mathematical Sciences at Seoul National University as a tenure-track faculty.

Mathematical theory of infinitely large neural networks in deep learning and GANs (90분)

As deep learning exhibits remarkable empirical success, there is a pressing need to establish a theoretical understanding. However, analyzing the practical training dynamics and establishing any theoretical guarantees for deep learning has been an intractable task. Therefore, the field of deep learning theory has started to study the dynamics of large neural networks in the limit of the network size being infinitely large. In other words, the infinitely large neural networks are studied as simplified models of the practical finite neural networks. In this talk, we will provide a high-level overview of the theory of infinitely large neural networks in the setup of deep supervised learning. We will then present some recent work on the theory of infinitely large networks in generative adversarial networks (GANs).