About Mingjing Xu

Snapshot

I am Mingjing Xu, a Computer Science PhD student at Swansea University since September 2025. My work sits around efficient and trustworthy AI: multi-objective optimization, multi-task learning, multimodal representation learning, and secure sensing systems.

Before Swansea, I worked across embedded software, electronic engineering, mobile systems, and machine learning. I started as an embedded software engineer at Jabil, moved into research through my M.Sc. at The Chinese University of Hong Kong, served as both a teaching assistant and research assistant at Temple University, and later conducted AI optimization research at Rochester Institute of Technology.

This page is a personal introduction rather than a formal PDF CV. For publication updates, the most reliable links are my Google Scholar, OpenReview, ORCID, and GitHub.

What I Work On

• Efficient optimization for AI: stochastic multi-objective optimization, multi-task learning, gradient manipulation, and learning algorithms that reduce unnecessary computation.
• Trustworthy and robust ML: adversarial behavior in sensing systems, privacy/security-aware AI, and robustness under real deployment constraints.
• Multimodal learning: representation learning across vision, language, tactile sensing, and edge/robotic contexts.
• Embedded and edge intelligence: practical systems that connect firmware, sensing hardware, and deployable machine learning.

Current Academic Path

• Sep 2025 - Present: PhD student in Computer Science, Faculty of Science and Engineering, Swansea University.
• Jan 2024 - Jan 2025: Research Assistant in Computing and Information Sciences, Rochester Institute of Technology, working on AI optimization, multi-objective learning, and multi-task learning.
• Jan 2023 - Dec 2023: Teaching Assistant and Research Assistant in Computer and Information Science, Temple University. Teaching class: CIS 3515, Introduction to Mobile Application Development. Research focus: trustworthy WiFi sensing and practical adversarial attacks on sensing systems.
• Aug 2021 - Nov 2022: M.Sc. in Electronic Engineering, The Chinese University of Hong Kong. Advisor: Prof. Hongliang Ren.
• Sep 2015 - Jul 2019: B.E. in Electrical Engineering and Automation, Shanghai University. Advisor: Prof. Zhiyuan Gao.

Research Highlights

• PSMGD, AAAI 2025, first author: this work studies multi-objective optimization for machine learning problems with multiple, potentially conflicting objectives. Existing gradient-manipulation methods often spend substantial time solving an auxiliary problem to compute dynamic objective weights at each step. PSMGD starts from the observation that these weights usually change slowly over short training intervals, so it periodically recomputes them and reuses them between updates. The result is a faster stochastic multi-gradient method with convergence guarantees for strongly convex, general convex, and non-convex objectives, plus a backpropagation-complexity view that better captures the practical cost of multi-objective training.
• TLV-CoRe, AAAI 2026, co-first author: this work targets tactile-language-vision alignment for robotic perception. Tactile sensing provides fine-grained physical information, but different sensors produce redundant and poorly standardized features. TLV-CoRe introduces a sensor-aware modulator, tactile-irrelevant decoupled learning, and a unified bridging adapter to build a shared representation space across tactile, language, and vision modalities. It also proposes an RSS evaluation view around robustness, synergy, and stability, making the paper both a method contribution and an evaluation contribution for multimodal tactile learning.
• Beyond Student / InherNet, ICLR 2026, co-first author: this work asks whether a compact model can inherit a teacher network more directly than the usual student-teacher distillation pipeline. InherNet uses asymmetric low-rank decomposition and SVD-based initialization to reconstruct a lightweight but expressive inheriting network from the teacher’s weights, aiming to preserve principal knowledge without heavy architectural disruption. The project connects efficient model compression, network inheritance, and transfer learning beyond standard knowledge distillation.
• WiFi sensing attack, ACM MobiCom 2024: this co-authored work was completed during my research period at Temple University. It shows that small, communication-level packet perturbations can create practical adversarial attacks against WiFi sensing, reflecting my broader interest in trustworthy AI and sensing systems where attacks may happen through the deployment substrate rather than through the model input alone. I was responsible for code design, conducting experiments, and writing the methods section of the paper.
• Origami robot, IEEE ICRA 2024: this collaboration contributed to a magnetic-guided flexible origami robot for long-term phototherapy of H. pylori in the stomach, connecting sensing, control, medical robotics, and deployable intelligent systems.

Publications

2026

• Y. Zhou, J. Shi, M. Xu (co-first author), Z. Jiang, and J. Chen, “Beyond Student: An Asymmetric Network for Neural Network Inheritance,” ICLR 2026 Poster. (CCF-A, CORE-A*)
• Y. Zhou, M. Xu (co-first author), J. Shi, Q. Li, and J. Chen, “Collaborative Representation Learning for Alignment of Tactile, Language, and Vision Modalities,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 22, pp. 18864-18872, 2026. (CCF-A, CORE-A*)
• W. Wang, M. Xu, Z. Cao, J. Guo, C. Liu, H. Zhang, and X. Zhang, “Unified Data Synthesis for Automated 3D Visual Inspection and Digital Twinning of Bridges,” Automation in Construction, vol. 182, article 106741, 2026. (JCR-Q1)

2025 And Earlier

• M. Xu, P. Ju, J. Liu, and H. Yang, “PSMGD: Periodic Stochastic Multi-Gradient Descent for Fast Multi-Objective Optimization,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 20, pp. 21770-21778, 2025. (CCF-A, CORE-A*)
• C. Li, M. Xu, Y. Du, L. Liu, C. Shi, Y. Wang, H. Liu, and Y. Chen, “Practical Adversarial Attack on WiFi Sensing Through Unnoticeable Communication Packet Perturbation,” Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pp. 373-387, 2024. (CCF-A, CORE-A*)
• S. Yuan, B. Liang, P. W. Wong, M. Xu, C. H. Li, Z. Li, and H. Ren, “Magnetic-Guided Flexible Origami Robot toward Long-Term Phototherapy of H. pylori in the Stomach,” 2024 IEEE International Conference on Robotics and Automation, pp. 9851-9857, 2024. (CCF-B, CORE-A*)
• Mingjing Xu, master’s thesis, “Modeling and Vision-Assisted Optimization of Sampling Strategies for Lesion Detection.”

Reviewing And Service

• ICML 2026 Gold Reviewer Award.
• Program Committee / reviewer service for major AI and machine learning venues, including AAAI, NeurIPS, ICML, and ICLR.
• My reviewing interests are closest to optimization, multi-task learning, robust/trustworthy ML, multimodal representation learning, and applied AI systems.

Selected Honors

• 2026 ICML Gold Reviewer Award.
• 2022 M.Sc. Certificate of Merit, CUHK Graduation Scholarships.
• 2018 First Prize, Service Robot Special Competition of the China Robot Competition.
• 2018 Second Prize, TI Cup Undergraduate Shanghai Electronics Design Contest.
• 2016 and 2017 National Encouragement Scholarship, Shanghai University.
• Selected certificates: DeepLearning.AI Deep Learning Specialization, Stanford Machine Learning, UMich Python for Everybody, Google Data Analytics, and PKU Python Language Foundation and Application.

Experience Before The PhD

Embedded Software At Jabil

From 2019 to 2021, I worked as an embedded software engineer at Jabil on IoT and 4G/5G radio-related products. The role was close to real hardware: firmware bring-up, board-level debugging, driver integration, communication-protocol testing, and cross-team issue tracking were part of the daily work.

Technically, this period gave me a practical foundation in MCU selection, peripheral interfaces, BSP and driver development, RT-Thread workflows, Xilinx/Vivado tooling, JSON-based configuration and testing utilities, Git/Azure DevOps collaboration, and hardware/software integration. It also taught me engineering habits that still affect my research work: isolate the failure mode first, respect measurement evidence, write reproducible notes, and avoid solutions that only work in a clean demo environment.

Looking back, Jabil was important because it made me comfortable with deployment constraints. Later, when I moved into AI and optimization research, I kept thinking about whether an algorithm is only mathematically elegant or also efficient enough to run under real compute, memory, and debugging constraints.

Teaching And Research At Temple

At Temple University, I served as both a teaching assistant and a research assistant in Computer and Information Science. As a TA for CIS 3515, an Android/Kotlin mobile application development course, I helped students understand mobile software design, debugging, and implementation. Teaching helped me turn scattered engineering experience into clearer explanations, and it made me much more deliberate about writing code that other people can read.

My research at Temple focused on trustworthy sensing and adversarial behavior in WiFi sensing systems. The ACM MobiCom 2024 paper, “Practical Adversarial Attack on WiFi Sensing Through Unnoticeable Communication Packet Perturbation,” came from this period. This work shaped my understanding of trustworthy AI in physical and wireless environments, especially cases where vulnerabilities arise from the sensing and communication substrate rather than from clean digital inputs.

Research At CUHK

At The Chinese University of Hong Kong, my research experience was rooted in electronic engineering, medical robotics, and vision-assisted optimization. My master’s thesis, “Modeling and Vision-Assisted Optimization of Sampling Strategies for Lesion Detection,” connected data-driven modeling with practical medical sensing and sampling decisions.

This stage trained me to think across modeling, sensing, control, and physical constraints. It also led to collaboration on magnetic-guided flexible origami robotics for long-term phototherapy of H. pylori in the stomach. Compared with purely software projects, this work made the gap between algorithm design and physical deployment very concrete.

AI Research At RIT

At Rochester Institute of Technology, my research direction shifted strongly toward machine learning theory and AI systems. I worked mainly on multi-objective optimization, multi-task learning, and efficient training methods, including the PSMGD project.

This period was where I began to connect my systems background with learning algorithms. PSMGD came from a practical concern that many multi-objective training methods are theoretically appealing but computationally expensive. At RIT, my work was mainly about making optimization and multi-task learning methods more efficient, principled, and practical for real AI training workloads.

Early Projects

• Built an embedded AI garbage-classification device with TensorFlow, RT-Thread, ART-Pi, camera input, screen output, and model deployment tooling. Repository: Garbage-Classification-Device.
• Led an STM32F1/RTOS-based loop-current signal detection device for the TI Cup Undergraduate Shanghai Electronics Design Contest.
• Led a service-robot project using MSP432, Bluetooth/MQTT communication, OpenMV tracking, and a WeChat mini program controller.

Skills

• Programming: Python, C/C++, Kotlin, Matlab, Git, Linux, LaTeX.
• AI and ML: multi-objective optimization, multi-task learning, multimodal learning, robustness, image/video analysis, experimental evaluation, and academic writing.
• Systems: embedded firmware, MCU drivers, RTOS workflows, IoT sensing, wireless systems, and hardware/software integration.
• Communication: academic reviewing, teaching, technical writing, and cross-disciplinary collaboration.

Outside Research

Outside work, I like table tennis, badminton, football, hiking, and mobile games. I am a long-time fan of Lionel Messi and FC Barcelona. I also enjoy natural scenery; my hometown is near Mount Langshan, a World Natural Heritage site that still shapes how I think about home.