Abstract

Deep learning, represented by deep artificial neural networks, has been dominating numerous important research fields in the past decade. Although the invention of the neural network was to mimic a human's brain, the current development of deep learning is not primarily driven by the increasingly growing understanding to the brain. Brain is the most intelligent system we have ever known so far, although the brain remains vastly undiscovered, it is clear that the existing deep learning still goes far behind human brain in many important aspects such as efficiency, interpretability, memory, etc. Given the incredible capability of the human brain, we argue that neuroscience can always offer support for deep learning as a think tank and a validation means. In this talk, we discuss drawing the mechanism of genome bottleneck into deep learning, with an emphasis on solving problems in model compression, which is to facilitate the deployment of large models.

Bio

Dr. Fenglei Fan is currently an Assistant Professor with Department of Data Science, City University of Hong Kong. He is the Chief Scientist of the Huawei Key Project. His primary research interests lie in NeuroAI and its applications in model compression. He was the recipients of the IBM AI Horizon Scholarship, the 2021 International Neural Network Society Doctoral Dissertation Award, and He won OlympusMons Pioneering Award, a prestigious award in the field of storage. He has one paper selected as one of few 2024 CVPR Best Paper Award Candidates, one won the IEEE Nuclear and Plasma Society IEEE TRPMS Best Paper Award, and one ESI highly cited paper. He organized special issues in journals like IEEE TRPMS, presented three tutorials in AAAI2023, IJCNN25, and WWW2025, and served as (senior) program committee members in AAAI and IJCAI.

Abstract
As an important medium for information propagation, the study of public opinion propagation mechanism on social networks can help people deeply understand the whole process of the outbreak, diffusion and extinction of public opinion events, s thereby significantly enhancing the governance ability of governments and enterprises against false or undesirable content. In the real world, the spread of public opinion events is profoundly influenced by the intricate public opinion environment. Within the complex public opinion environment, multifaceted factors such as user social relationships, individual preferences, characteristics of public opinion events, pivotal users, sphere effects, and various complex public opinion field effects collectively affect the propagation process of public opinion events. In this talk, we introduce recent studies on predicting the public opinion event propagation of large social networks under the complex public opinion field environment. These studies mainly encompass three aspects: inference of users' propagation preference, propagation model of public opinion events based on the sphere effect, and analysis and prediction of the propagation of public opinion events under the effect of complex public opinion field. These studies can achieve accurate prediction of public opinion propagation in complex public opinion field environments.

Bio
Dr. Yajun Yang is an Associate Professor and doctoral supervisor at the College of Intelligence and Computing, Tianjin University. He received his Ph.D. from the Harbin Institute of Technology and has held visiting scholar positions at The Chinese University of Hong Kong and Griffith University. His research interests focus on graph databases, graph mining, and graph learning. Dr. Yang has served as the principal investigator on multiple research projects funded by the National Natural Science Foundation of China (NSFC) and the National Key Research and Development Program of China. He has published more than 40 papers in leading international conferences and journals, including SIGMOD, VLDB, ICDE, KDD, NeurIPS, TOIS, AAAI, ICDM, CIKM, and DASFAA. He has also served as a program committee member for several international conferences, such as KDD, ICDE, CIKM, DASFAA, APWeb-WAIM, and ADMA.

Abstract
This report summarizes the recent advances about worker selection across crowdsourcing, social networks, and federated learning. In social diffusion, it introduces MT-DM, a multi-task diffusion incentive mechanism that fuses reverse auctions with Multi-Task RR sampling to maximize expected task quality under budget with truthfulness, individual rationality, and near-optimality, achieving major runtime gains. It then presents DQNSelector, which learns dual embeddings of long-range social influence and local sensing coverage with Rainbow DQN to choose seed workers, boosting effective coverage by 6–18% with sub-second inference. In federated learning, client selection with unknown data quality is modeled as a CMAB with a Stackelberg game, leveraging gradient-discrepancy estimation, CUCB exploration, and incentive design to attain a unique equilibrium, lower regret, and reduce training cost while preserving accuracy. Finally, a training-potential–driven joint client–modality selection framework mitigates multi-modal FL communication by 25–75% with negligible accuracy loss, validated by convergence analysis and extensive experiments.

Bio
Jianxiong Guo is an associate professor with the Advanced Institute of Natural Science, Beijing Normal University, and vice director of the Engineering Research Center of Cloud-Edge Intelligent Collaboration on Big Data, Ministry of Education. He received his Ph.D. from the Department of Computer Science, University of Texas at Dallas, in 2021, supervised by Dr. Ding-Zhu Du. He is mainly engaged in the research of combinatorial optimization, game and incentive mechanism design in social networks, distributed machine learning, Internet of Things, and cloud-edge collaborative systems, recently focusing on build efficient networking systems for AI applications. In recent years, he has published more than 100 academic papers in computer networks, data science, and theoretical computer science, among which more than 40 papers have been published in IEEE/ACM Transactions series as the first/corresponding author. He presided over or participated in several national projects, including NSFC and National Key R&D program of China, and served as a member of academic committees of many top international conferences and a reviewer of international journals.

Abstract

Image restoration is a fundamental task in computer vision, with photo restoration drawing increasing attention due to its high practical value. This report first highlights the key challenges in photo restoration, including handling scratched regions, developing discriminative models for diverse degradation types, and the limited availability of annotated datasets. To address these challenges, we construct dedicated datasets for photo restoration and propose several solutions: a contextual-assisted scratched photo restoration method, a hierarchical damage correlation framework for old photo restoration, and a semi-supervised approach that leverages important samples for old photo restoration. Moreover, for underwater image scenarios, we present a novel image quality assessment method that combines physical priors with human perceptual factors. Extensive quantitative and qualitative experiments verify the effectiveness and superiority of the proposed methods.

Bio

Weiwei Cai is currently a Postdoctoral Fellow at The Hong Kong Polytechnic University. She obtained her Ph.D. degree in Computer Science and Engineering from South China University of Technology in 2024. Her research focuses on computer vision, image processing, computer graphics, and deep learning, with a particular emphasis on developing advanced algorithms for visual understanding and intelligent image analysis. Dr. Cai has published her work in prestigious journals and conferences, including IEEE TCSVT, IEEE TMM, IEEE TIP, Information Fusion, IJCAI, and ICASSP. From September 2025 to September 2027, she is supported by the Hong Kong Postdoctoral Matching Scheme. She actively contributes to the academic community by serving as a reviewer for leading journals and conferences, such as IEEE TCSVT, Pattern Recognition, CVPR, ECCV, and ICASSP.