Invited Speaker
Invited Talk: Bayesian Optimization over Combinatorial Spaces
Seungjin Choi
Director of Machine Learning LabIntellicode, Korea
Biography:
Seungjin Choi received his B.S. and M.S. degrees in electrical engineering from Seoul National University, Korea in 1987 and 1989, respectively, and his Ph.D. in electrical engineering from the University of Notre Dame, Indiana, in 1996. In 1997, he worked as a Frontier researcher at the Laboratory for Artificial Brain Systems, RIKEN, Japan, where he focused on independent component analysis under the guidance of Prof. Andrzej Cichocki and Prof. Shunichi Amari. From 1997 to 2000, he was an Assistant Professor in the School of Electrical and Electronics Engineering, Chungbuk National University. Between 2001 and 2019, he was a Professor of Computer Science at Pohang University of Science and Technology, Korea. He had held advisory roles for the Shinhan Card Bigdata Center, Samsung Research, and the Samsung Advanced Institute of Technology. Transitioning to industry, he served as the Chief Technology Officer (CTO) at BARO AI and as an executive advisor for BARO AI Academy, where he developed online lectures for machine learning, deep learning, and mathematics for machine learning. Since 2022, he has been with Intellicode as the director of the Machine Learning Lab, leading research and developing algorithms for customized AI solutions tailored to various business needs. He also served as president of the AI Society within the Korea Institute of Information Scientists and Engineers. Seungjin Choi has contributed to the machine learning and AI community as an area chair for top-tier conferences, including NeurIPS, ICML, ICLR, AISTATS, AAAI and IJCAI. His current research interests include Bayesian optimization, uncertainty quantification, probabilistic inference, conformal inference, and causal inference.
Abstract:
Bayesian optimization is a powerful tool for optimizing expensive black-box functions, commonly used in scenarios where function evaluations are costly or time-consuming. Its applications span a wide range, including hyperparameter optimization, automated machine learning, protein/DNA sequence design, material design, portfolio optimization, and so on. The standard Bayesian optimization leverages Gaussian process regression for constructing a surrogate model to estimate a landscape of the black-box objective, assuming decision variables are continuous in Euclidean space. While much of the traditional focus has been on continuous optimization, many real-world problems require optimizing over combinatorial spaces, introducing unique challenges. In this talk, I will begin with a brief overview of standard Bayesian optimization, highlighting its core principles. Then I will delve into recent advancements in combinatorial Bayesian optimization, presenting key methodologies tailored to handle the discrete and often high-dimensional nature of these spaces. A special focus will be on my own work leveraging random projection techniques to efficiently optimize within combinatorial domains. Finally I will discuss an emerging trend of utilizing large language models (LLMs) as surrogate models in Bayesian optimization, illustrating their potential to reshape the field of AI-assisted drug discovery.
Keynote I: Benchmark Datasets to Evaluate Safety and Faithfulness of Generative AI
Ho-Jin Choi
ProfessorSchool of Computing
Korea Advanced Institute of Science and Technology (KAIST)
Biography:
Dr. Ho-Jin Choi is currently a professor in the School of Computing at KAIST, Daejeon, Korea. In 1982, he received a BS in Computer Engineering from Seoul National University, Korea, in 1985, an MSc in Computing Software and Systems Design from Newcastle University, UK, and in 1995, a PhD in Artificial Intelligence from Imperial College London, UK. From 1982 to 1989, he worked as a senior engineer for Data Communications Co. of Korea (DACOM), Seoul, Korea, and between 1995 and 1996, as a post-doctoral researcher at IC-PARC, Imperial College London. From 1997 to 2002, he worked as an assistant professor in the School of Electronics, Information Communication, and Computer Engineering at Korea Aerospace University, Korea, then from 2002 he has been with KAIST. Between 2002 and 2003, he visited Carnegie Mellon University (CMU), Pittsburgh, USA, then served as an adjunct faculty for CMU Master of Software Engineering (MSE) program operated by the Institute for Software Research (ISR). Between 2006 and 2008, he served as the Director of Institute for IT Gifted Youth. Between 2019 and 2022, he served as the Chief of KAIST Convergence AMP and as the Chief of Software Graduate Program. Since 2018, he has served as the Director of Smart Energy Artificial Intelligence Research Center, and since 2020, as the Director of Center for Artificial Intelligence Research (CAIR). Over the past 20 years, he served for the boards of directors for several academic communities such as the AI Society, the Software Engineering Society, and the Language Technology SIG within the Korean Institute of Information Scientists and Engineers (KIISE), and also for the Korean Society of Medical Informatics.
Abstract:
Generative AI techniques have opened new business opportunities that can change and impact our ordinary lives. However, evaluating and measuring the safety and faithfulness of the AI-generated contents remains generally under-explored. This talk introduces some research efforts to construct benchmark datasets for evaluating safety and faithfulness of generative AI, currently undergoing in South Korea by TTA (Telecommunications Technology Association) with the collaboration of researchers and practitioners from KAIST, University of Seoul, Keimyung University, SelectStar, and Kakao. Our ultimate goal is to establish a methodology for constructing benchmark datasets, evaluation metrics, protocols and annotations to evaluate the safety and faithfulness of multimodal generative AI.
Keynote II: A progress on design and analysis of Adversarial Neural Cryptosystems after Abadi-Anderson@GoogleBrain[2016]
Kouichi Sakurai
ProfessorCyber Security Center （Concurrent)
School of Engineering Department of Electrical Engineering and Computer Science (Concurrent)
Biography:Kouichi Sakurai received the B.S. degree in mathematics from the Faculty of Science, Kyushu University in 1986. He received the M.S. degree in applied science in 1988, and the Doctorate in engineering in 1993 from the Faculty of Engineering, Kyushu University. He was engaged in research and development on cryptography and information security at the Computer and Information Systems Laboratory at Mitsubishi Electric Corporation from 1988 to 1994. From 1994, he worked for the Dept. of Computer Science of Kyushu University in the capacity of associate professor and became a full professor there in 2002. He had been working also with the Institute of Systems & Information Technologies and Nanotechnologies, as the chief of Information Security laboratory, for promoting research cooperation among the industry, university and government under the theme "Enhancing IT-security in social systems". He has been successful in generating such co-operation between Japan, China and Korea for security technologies as the leader of a Cooperative International Research Project supported by the National Institute of Information and Communications Technology (NICT) during 2005-2006. Moreover, in March 2006, he established research cooperation under a Memorandum of Understanding in the field of information security with Professor Bimal Kumar Roy, the first time Japan has partnered with The Cryptology Research Society of India (CRSI). Professor Sakurai has published more than 450 academic papers around cryptography and information security (See: https://dblp.org/pid/16/3865.html ).
Abstract:
In 2016, Abadi and Andersen of the Google Brain lab. presented a model in which adversarial machine learning was applied to the construction of cryptographic communications. Here, the design and security evaluation of the cipher is performed entirely by artificial intelligence alone, without any human intervention; the results of Abadi et al.'s work showed that machines can generate ciphers while simultaneously performing design and analysis, opening up the field of cryptography and deep learning integration. However, Abadi et al.'s primitive model still faces research challenges, such as strengthening the generated ciphers, extending them to multi-party group communication, and achieving practicality and flexibility in replacing conventional cryptography, such as realising public key cryptography that can be used by an unspecified number of people. This talk gives a survey on research progress of Neural Cryptography: before vs. after the Google-2016, and introduces recent results by the speaker’s research group.
