Speakers

Prof. James Tin-Yau KWOK, IEEE Fellow

 Hong Kong University of Science and Technology

James Kwok is a Professor in the Department of Computer Science and Engineering, Hong Kong University of Science and Technology. He is an IEEE Fellow.  He has served / is serving as an Associate Editor for the IEEE Transactions on Neural Networks and Learning Systems, Neural Networks, Neurocomputing, Artificial Intelligence Journal, International Journal of Data Science and Analytics, and on the Editorial Board of Machine Learning. He is also serving as Senior Area Chairs of major machine learning / AI conferences including NeurIPS, ICML, ICLR, IJCAI.  He is on the IJCAI Board of Trustees. He is recognized as the Most Influential Scholar Award Honorable Mention for "outstanding and vibrant contributions to the field of AAAI/IJCAI between 2009 and 2019". Prof Kwok will be the IJCAI-2025 Program Chair.

Title: Enhancing Language Models through Improved Pre-Training and Fine-Tuning

Abstract: Language models (LMs) are essential in natural language processing and vision-language modeling. However, several challenges arise in pre-training and fine-tuning of LMs. First, when learning through unsupervised pre-training, information that are semantically irrelevant may negatively affect downstream tasks, leading to negative transfer. Second, cross-modal masked language modeling is often used to learn vision-language associations in vision-language models. However, existing masking strategies may be insufficient in that the masked tokens can sometimes be simply recovered with only the language information, ignoring the visual inputs. Lastly, prompt tuning is effective in fine-tuning LMs on downstream tasks with limited labeled samples, but prompt design is difficult.  

To tackle these issues, we propose several measures. First, we introduce a new pre-training method that trains each expert with only semantically relevant data through cluster-conditional gates. This allows downstream tasks be allocated to customized models pre-trained on data most similar to the downstream data. Second, on pre-training vision-language models, we use a masking strategy based on the saliencies of language tokens to the image. Lastly, we use meta-learning to learn an efficient prompt pool that can extract diverse knowledge from historical tasks. This allows instance-dependent prompts to be constructed from the pool without tuning the whole LM. Experimental results show that these measures can significantly improve the performance of LMs. 

Prof. Zuqing Zhu 

IEEE Fellow, AAIA Fellow

University of Science and Technology of China

Zuqing Zhu is a full professor at the School of Information Science and Technology, University of Science and Technology of China (USTC). Since January 2011, he has been leading the Intelligent Networking System and Future InterNet Infrastructure (INFINITE) Lab at USTC. Dr. Zhu is currently the Chair of the Technical Committee on Optical Networking (ONTC) of the IEEE Communications Society (ComSoc). He has been the Lead Series Editor of the Optical Communications and Networks Series in IEEE Communications Magazine and is currently on the editorial boards of IEEE Transactions on Network and Service Management, Optics Express, IEEE Transactions on Communications, Optical Switching and Networking (Elsevier), etc. Dr. Zhu is the Steering Committee Chair of the IEEE International Conference on High Performance Switching and Routing (HPSR), a Member-at-Large of the GLOBECOM/ICC Technical Content (GITC) Committee, and has been an IEEE ComSoc Distinguished Lecturer (2018–2021). He has published more than 360 papers in referred journals and conferences of IEEE and Optica. His h-index is 57 (from Google Scholar, 12/2023), his Google Scholar profile is here, and his ResearchGate profile can be found here. He has received the 2023 Distinguished Technical Achievement Award from the Communications Switching and Routing Technical Committee (CSR-TC) of ComSoc. Dr. Zhu is a Fellow of IEEE, a Fellow of the Asia-Pacific Artificial Intelligence Association (AAIA), and a Senior Member of Optica (formerly OSA).

Title: Machine Learning in and for Optical Data-Center Networks

Abstract: In the first part of this talk, we will first discuss the challenges on scalability, energy and manageability of data-center network (DCN) systems, and then explain why all-optical inter-connection can be a promising solution for future DCN systems. Next, we describe a novel all-optical inter-connection architecture based on arrayed waveguide grating router (AWGR) and wavelength-selective switches (WSS'), namely, Hyper-FleX-LION, explain its operation principle, and show experimental results of running distributed machine learning (DML) in a DCN in Hyper-FleX-LION. In the second part of this talk, we will explain how machine learning can be leveraged to realized knowledge-defined networking (KDN) and facilitate network automation in DCNs. Experimental results demonstrate that KDN can automatically reduce task completion time.

Prof. Qing Li, IEEE Fellow, IET/IEE Fellow

The Hong Kong Polytechnic University

Qing Li is a Chair Professor and Head of the Department of Computing, the Hong Kong Polytechnic University. He received his B.Eng. from Hunan University (Changsha), and M.Sc. and Ph.D. degrees from the University of Southern California (Los Angeles), all in computer science. His research interests include multi-modal data management, conceptual data modeling, social media, Web services, and e-learning systems. He has authored/co-authored over 500 publications in these areas, with over 37000 citations and H-index of 79 (source: Google Scholars). He is actively involved in the research community and has served as Editor-in-Chief of Computer & Education: X Realitty (CEXR) by Elsevier, associate editor of IEEE Transactions on Artificial Intelligence (TAI), IEEE Transactions on Cognitive and Developmental Systems (TCDS), IEEE Transactions on Knowledge and Data Engineering (TKDE), ACM Transactions on Internet Technology (TOIT), Data Science and Engineering (DSE), and World Wide Web (WWW) Journal, in addition to being a Conference and Program Chair/Co-Chair of numerous major international conferences. He also sits/sat in the Steering Committees of DASFAA, ER, ACM RecSys, IEEE U-MEDIA, and ICWL. Prof. Li is a Fellow of IEEE, AAIA, and IET.

Title: Toward an Edu-Metaverse Supporting Immersive Explorations and Collaborative Learning Through Knowledge Graph and VR Techniques

Abstract: Metaverse as an education platform aims at bringing students and educators together into an interactive virtual environment that could potentially unleash a much richer educational content medium due to the highly immersive learning experience. The driving forces railing the development of engaging education interactions between instructors and students in a metaverse environment stem from (1) the need to expand educational access, and (2) enhancing the convenience of learning processes. First, knowledge graphs (KGs) are increasingly been built for pedagogical purposes. To depict the rich but latent relations among different concepts in a course textbook, course KGs are constructed and refined interactively. However, the application of course KGs for real study scenarios and student career development remains largely unexplored and nontrivial.  In this talk, we present a novel tool exploiting course knowledge graphs, to facilitate both intra-course study and inter-course development for students significantly. An interactive web system has been developed for both instructors to construct and manipulate course KGs, and for students to view and interact with knowledge concepts. Next, to visualize the centrality of a course KG based on various metrics, concept-level advising is designed, through which we propose a tailored algorithm to suggest the learning path based on what concepts students have learned. Course-level advising is instantiated with a course network, which indicates the prerequisite relations among different levels of courses, corresponding to the annually increasing curricular design and forming different major streams. Through building such an edu-metaverse, our work solves a pressing issue for edu-metaverse on how it can manifest to connect a broad range of learning material and educational concept together on a ubiquitous platform for users to learn and explore knowledge. To facilitate association, exploration, and engagement in collaborative learning, we combine the structure of KGs and the immersion of virtual reality (VR) in our pilot metaverse prototype, K-Cube VR, which is developed and tested to validate the underlying edu-metaverse theory and framework. Examples will also be provided to illustrate the effectiveness of our Edu-Metaverse approach.

Prof. Pietro S. Oliveto

Southern University of Science and Technology

Pietro Oliveto holds a Laurea degree in computer science from the University of Catania, Italy, awarded in 2005, and a PhD degree from the University of Birmingham, UK, conferred in 2009. His academic journey has been marked by several prestigious fellowships, including the EPSRC PhD+ Fellowship (2009-2010) and EPSRC Postdoctoral Fellowship (2010-2013) at the University of Birmingham, followed by the Vice-Chancellor's Fellowship (2013-2016) and EPSRC Early Career Fellowship (2015-2020) at the University of Sheffield. Prior to joining SUSTech, he served as the Chair in Algorithms at the Department of Computer Science, University of Sheffield.

Professor Oliveto's primary research focus is on the performance analysis, particularly the time complexity, of bio-inspired computation techniques. These techniques include evolutionary algorithms, genetic programming, artificial immune systems, hyper-heuristics, and algorithm configuration. Currently, he is spearheading the establishment of a Theory of Artificial Intelligence Lab at SUSTech.

His contributions to the academic community extend beyond research, as he has guest-edited special issues for journals such as Computer Science and Technology, Evolutionary Computation, Theoretical Computer Science, IEEE Transactions on Evolutionary Computation, and Algorithmica. He has also co-chaired the IEEE symposium on Foundations of Computational Intelligence (FOCI) from 2015 to 2021 and served as co-program Chair for the ACM Conference on Foundations of Genetic Algorithms (FOGA 2021). Additionally, he has held the position of Theory Track co-chair at GECCO 2022 and GECCO 2023. Professor Oliveto is a member of the Steering Committee of the annual workshop on Theory of Randomized Search Heuristics (ThRaSH), served as the Leader of the Benchmarking Working Group for the COST Action ImAppNIO, is a member of the EPSRC Peer Review College, and serves as an Associate Editor for IEEE Transactions on Evolutionary Computation.

Title: Computational Complexity Analysis of Sexual Evolution for the Design of Better General Purpose Algorithms for AI

Abstract: Large classes of the general-purpose optimisation algorithms at the heart of modern artificial intelligence and machine learning technologies are inspired by models of Darwinian evolution. In this talk we show how the foundational computational complexity analysis of such algorithms leads to an understanding of their behaviour and performance. Such understanding in turn allows informed decisions on how to set their many parameters and how to improve the algorithms to allow for the obtainment of better solutions in shorter time. We provide two concrete examples of how such analyses can lead to counter intuitive insights into how to design sexual evolution inspired algorithms (using populations and recombination) and how to set their parameters such that they can considerably outperform their single trajectory and mutation only (asexual) counterparts at hillclimbing unimodal functions, and at escaping from local optima. We conclude the talk by presenting experimental results that confirm the superiority of the designed algorithms that was proven for benchmark functions with significant structures, for classical combinatorial optimisation problems with practical applications.

Prof. Hao Xu

Jilin University

Xu Hao, professor and PhD supervisor at College of Computer Science and Technology, Jilin University, courtesy professor of University of Trento, Italy. Prof. Xu obtained his PhD degree from Department of Computer Science and Information Engineering, University of Trento in 2012, and completed his master and bachelor from College of Computer Science and Technology, Jilin University, China.

Professor Xu's research interests encompass knowledge graph, multimodal semantic understanding, and Human-Centered AI, as well as the interdisciplinary research on AI for education, health, and culture heritage. He has coordinated over 20 projects funded by National Natural Science Foundation of China, EU and industries. Prof. Xu has published over 90 research papers in international conferences and journals, including CVPR, ACL, IJCAI, and AAAI. He has been honored with the 2023 ACM UbiComP Distinguished Paper Award, the 2024 China Invention and Entrepreneurship Award, and the First Prize for National Business Science and Technology Progress in 2020 and 2015 respectively.

Title: Towards Human-Centered Artificial Intelligence

Prof. Xibin Wang

Jilin University

WANG Xibin obtained his B.S. degree in Electronic Science and Technology, M.S., and Ph.D. degree in Microelectronics and Solid Electronics from Jilin University, Changchun, China, in 2008, 2010, and 2013, respectively. He is currently a professor in Jilin University and the State Key Laboratory of Integrated Optoelectronics. In 2016, he was selected into the "Hong Kong Scholars" program and was named a "Hong Kong Scholar". From 2017 to 2018, he was engaged in post-doctoral research at the City University of Hong Kong. In 2018, he was selected into the outstanding young teacher training program of Jilin University. His research field focuses on polymer and Si waveguide integrated devices for optical communication, optical interconnections, as well as optical sensing. He has published over 100 SCI retrieval papers in international journals and authorized 21 China’s national invention patents, three of which have been transferred to enterprises.

Title: Multimode optical switch based on planar waveguides

Abstract: Multimode optical switch is becoming more and more important as one of the key elements for mode-division multiplexing (MDM) optical communication systems. However, the different structural relationships between the waveguides and the different modes make it challenging to control more spatial modes simultaneously. Here, we put forward some multimode thermo-optic (TO) switch based on planar waveguide. The proposed device can realize the switching function for the guided four modes insensitively. Our fabricated device based on polymer waveguide platform shows a low power-consumption. The proposed MWOS can be also scaled to accommodate more spatial modes flexibly and easily, which can serve as an important building block for MDM systems.

Assoc. Prof. Pavel Loskot

Zhejiang University-University of Illinois at Urbana-Champaign Institute (ZJUI)

Pavel Loskot joined the ZJU-UIUC Institute in January 2021 as the Associate Professor after being nearly 14 years with Swansea University in the UK. He received his PhD degree in Wireless Communications from the University of Alberta in Canada, and the MSc and BSc degrees in Radioelectronics and Biomedical Electronics, respectively, from the Czech Technical University of Prague in the Czech Republic. He is the Senior Member of the IEEE, Fellow of the Higher Education Academy in the UK, and the Recognized Research Supervisor of the UK Council for Graduate Education. His current research interest focuses on problems involving statistical signal processing and importing methods from Telecommunication Engineering and Computer Science to other disciplines in order to improve the efficiency and the information power of system modeling and analysis.

Title: Kolmogorov-Arnold Network: A New Building Block of Deep Neural Networks

Abstract: The key idea behind deep neural networks (DNNs) is to provide a sufficiently rich model structure, which can be universally used to describe different datasets. The basic building block of DNNs is the multilayer perceptron having learnable weights with fixed non-linearity at its output. The Kolmogorov-Arnold networks (KANs) exploit learnable non-linearities that are used to transform data as they traverse across model branches. This enables much more flexible, and thus, potentially much more enhanced learning across a wider range of datasets. In this talk, I will discuss the learning opportunities and other benefits of KANs, but I will also mention some potential problems that may arise due to a greatly increased flexibility of KANs.