Professor Mingsen Deng
Guizhou University of Finance and Economics, China
Dr.Mingsen Deng is a Professor in the School of Information, Guizhou University of Finance and Economics, China. He serves as the Director of the Guizhou Provincial Key Laboratory of Computing and Network Convergence.Recognized as a Provincial-level Expert of Guizhou and a recipient of the "Hundred-level" High-level Innovative Talent Program of Guizhou, he has been honored with the Guizhou Provincial Government Special Allowance, the Guizhou Youth Science and Technology Award, and the Outstanding Science and Technology Worker award from the Chinese Institute of Electronics. As the head of the Guizhou Provincial First-class Discipline for Computer Science and Technology and a leader of a provincial-level technological innovation talent team, Dr. Deng has long focused his research on distributed parallel computing, electronic structure calculation algorithms, and data characteristics and management. He has successfully presided over more than 20 projects, including those funded by the National Natural Science Foundation of China, central government projects guiding local scientific and technological development, and major provincial-level research achievement transformation projects. He has published over 100 papers in mainstream international SCI journals such as IEEE Transactions, holds more than 40 domestic invention patents and one U.S. invention patent, and participated in the drafting of three current national standards. For his achievements, he has received numerous prestigious awards as the lead investigator, including the second prize of the Guizhou Provincial Natural Science Award and the first prize of the Provincial Teaching Achievement Award. Furthermore, he has supervised over 70 doctoral and master's graduates, many of whom have been selected for provincial and ministerial-level talent programs.
Title: Intelligent Task Scheduling under Complex Constraints in Computing Power Networks: From Reinforcement Learning to Causal Inference
Abstract: Computing power networks and air-ground collaborative computing have become critical infrastructures supporting applications with low-latency and high-computing demands. However, in highly dynamic network environments, task scheduling not only faces the trade-offs of multi-dimensional conflicting objectives but is also restricted by extremely complex system constraints. Addressing the challenges of time-varying user preferences and multi-objective conflicts in dynamic networks, we integrated deep reinforcement learning with evolutionary algorithms and effectively solved the high-dimensional policy exploration problem in scenarios featuring service caching dependencies and UAV-assisted air-ground collaboration, achieving a Pareto equilibrium in system performance. With the deepening of computing and network convergence, system constraints exhibit causal coupling characteristics, meaning that satisfying one constraint (e.g., latency) often cascade-triggers violations of other constraints (e.g., computing capacity, bandwidth, and energy consumption). Because traditional algorithms treat constraints as independent and equal penalty terms, they easily fall into local traps within large infeasible regions during the optimization process. We proposed a constraint causal learning method based on counterfactual reasoning. By mining the causal propagation topology of constraint violations from the evolutionary history, it accurately identified the root-cause constraints that trigger cascading failures, and used this to guide adaptive constraint relaxation and mutation operations. Consequently, it achieves highly efficient convergence in extremely narrow and fragmented feasible regions, demonstrating the evolution of scheduling algorithms in computing power networks from data-driven reinforcement learning to mechanism-driven causal inference.
Professor Haisheng Yu
China Internet Network Information Center, China
HaishengYu, Ph.D. Deputy Director of the Industry Development Department, China Internet Network Information Center (CNNIC), Member of the Public Security Communications Committee, China Institute of Communications Senior Member of IEEE Senior Member of China Computer Federation (CCF) Chair of IEEE Blockchain Association Macao & Guangzhou Chapters. He previously served as Director of the Forward-looking Technology Laboratory at the China Future Internet Engineering Center (CFIEC), and Visiting Research Fellow at the Macau University of Science and Technology (holding a Macao Blue Card). Dr. Yu has published more than 30 SCI and EI papers, including multiple papers in Journal Citation Reports (JCR) Q1 journals, and holds over 20 patents. He acted as General Co-Chair of IEEE Global Blockchain Congress 2025, and received the Best Paper Award of ICCIC 2025. He has served as a Program Committee Member for numerous international conferences including Trustcom, ICA3PP, MetaCom, UIC, NSS and BDCloud. He has led and participated in multiple international standards developed by ETSI, IETF and IEEE, as well as 10 national standards for IPv6. His research achievements have won the Second Prize of Science and Technology Progress Award from the China Institute of Communications and the First Prize of Science and Technology Progress Award from the China General Chamber of Commerce respectively.
Title: Research and Reflections on Agent Interconnection Networks
Abstract: This paper conducts research on the Agent network support system for cross-domain collaboration of AI Agents. Aiming at five core pain points in current A2A communications, namely the lack of Agent identity identification at the network layer, uncontrollable path scheduling, broken audit trails, disjointed registration, addressing and authorization mechanisms, and privacy and over-authorization risks brought by highly autonomous Agents, it proposes a trusted interconnection architecture that incrementally reuses existing Internet standards including IPv6, SRv6 and RDAP as well as audit nodes. The architecture consists of four collaborative planes: Agent Initiation, Registration-Discovery-Addressing, Trusted Bearer, and Policy-Audit-Traceback. Supported by four core technical solutions—the unified Agent-ID identity system, lightweight identity metadata via IPv6 extension headers, deterministic hierarchical path orchestration based on SRv6, and traceable governance through hierarchical audit nodes—the system realizes identifiable identities, orchestratable paths, auditable invocations and end-to-end trusted control for AI Agent communications. It follows the principle of incremental deployment without replacing existing Internet protocols and maintains compatibility with regular IPv6 communications.
Professor Huihuang Zhao
Hengyang Normal University, China
Huihuang Zhao, Dean of the School of Computer Science and Technology, Hengyang Normal University, Ph.D., Professor, Master's Supervisor. He is a postdoctoral fellow from Cardiff University (UK) and a researcher at the National Engineering Laboratory for Robot Visual Perception and Control Technology at Hunan University.Prof. Zhao completed his postdoctoral research at Hunan University (supervised by Academician Yaonan Wang) and served as a visiting scholar at the University of Texas Rio Grande Valley. He is recognized as a Young Backbone Teacher in Hunan Province and was selected for the Hengyang "Young Scientific and Technological Talent" Support Program. As an IEEE member, he serves as a director for both the Hunan Association for Artificial Intelligence and the Hunan Computer Society. He is a recipient of support from the Hengyang Normal University "Elite Talent Program," an Outstanding Graduate Supervisor of Hunan Province, and a nominee for the 18th Hengyang "Youth May Fourth Medal." Prof. Zhao has presided over numerous research projects, including two National Natural Science Foundation of China projects and one Ministry of Education project. He has published over 60 academic papers, with more than 50 indexed by SCI/EI, and has applied for 6 invention patents and over 20 software copyrights.
Title: Image generation technology and application based on style transfer
Abstract: Style transfer technology, as an important branch in the field of image generation, aims to transfer the style features of reference images to target content images while maintaining their semantic structure unchanged. This report systematically outlines the development trajectory and cutting-edge trends of style transfer technology. Firstly, from the perspective of methodological evolution, the limitations of traditional non realistic rendering methods were reviewed, with a focus on analyzing the three major technical paradigms based on deep learning: iterative optimization methods represented by Gatys et al., perceptual loss and feedforward network methods represented by Johnson et al., and data-driven methods based on generative adversarial networks and diffusion models. The inherent trade-offs between each paradigm in terms of generation quality, computational efficiency, and style controllability were revealed. Secondly, from the perspective of technological deepening, the extension of style transfer capability boundaries based on adaptive instance normalization, attention mechanism, Transformer architecture, and large-scale visual language models was explored, especially in emerging directions such as diffusion model style transfer and personalized customization generation. Finally, some achievements of our team in the field of style transfer based image generation technology were introduced.
ProfessorXiaoping Wu
Kaili University, China
XiaopingWu, Doctor of Engineering, Professor. Visiting Scholar under the "XiBuZhiGuang" Program of the Chinese Academy of Sciences. Director of the Engineering Research Center of Ministry of Education for Micro-Nano and Intelligent Manufacturing. Director of the Key Laboratory of Glass in Qiandongnan Prefecture. Dean of the School of Microelectronics and Artificial Intelligence, Kaili University.National Correspondent Expert of ISO/IEC JTC1/SC6. Council Member of the Educational Big Data Committee of the Chinese Education Development Strategy Society. Council Member of the Integrated Circuit Society of Guizhou Province. Council Member of the Artificial Intelligence Society of Guizhou Province. Member of the Scientific and Technological Innovation Talent Team for Communication and Information System of Guizhou Province.His main research field is Natural Language Processing (NLP). He has presided over 32 teaching and research projects with a total funding of more than 7.5 million yuan, and guided enterprises to invest more than 6.4 million yuan. He has published 45 academic papers, formulated 11 standards, authorized 3 patents, and developed 1 Provincial First-Class Course.
Title: Acoustic Feature Extraction and Speech Recognition Modeling for Low-Resource Languages
Abstract: This study takes low-resource languages represented by the Hmong language as the research object, with a focus on analyzing acoustic feature extraction and speech recognition for such languages. It aims to realize the effective preservation and intelligent application of ethnic minority languages through technological innovation.Three core challenges exist in this research: scarce linguistic data, complex acoustic characteristics, and obstacles in semantic comprehension. This paper proposes a systematic solution: steady-state signals are obtained via speech framing, windowing and short-time transformation, and Mel filter bank cepstral mapping is adopted to simulate human auditory perception features. CMVN is utilized to smooth spectral variations caused by different speakers and diverse background noises, regularizing speech signals to accelerate the convergence of acoustic models and enhance their robustness. Meanwhile, tone recognition and phoneme classification are jointly optimized to guarantee the precise capture of phonological features unique to the Hmong language. Drawing on the reference WeNet model, this work explores a hybrid model combining the Conformer-based end-to-end architecture with Connectionist Temporal Classification (CTC).Experimental results demonstrate that the proposed method achieves a low word error rate (WER) and strong generalization performance, which verifies its suitability for recognition tasks targeting low-resource languages like Hmong.