Projects

  • Multimodal Question Answering [Project Website]

    • Analyzed and visualized limitations SOTA Visual Question Answering(VQA) datasets in Python including VQA-2.0, GQA and TallyQA, and created an unbiased VQA dataset.
    • Explored and compared different SOTA cross-modality models on AWS EC2 including ViLT and LXMERT, and ran tests on current VQA benchmarks to evaluate the shortcomings of of the models.
    • Made in-depth error analysis on LXMERT and ViLT’s performance, and generated specific types of questions that existing baselines have weak performance in after introducing scene graph generation module and instance mask generation module to the baselines.
    • Designed an end-to-end question answering system that can utilize auxiliary tasks including instance segmentation and scene graph generation to learn better representations in images based on existing baseline models.

  • Tracking Semantic Evolutionary Changes in Large-Scale Medical Knowledge Bases

    • Developed an advanced reasoning approach to tracking the semantic difference in the meanings of medical terms between different versions of medical knowledge bases.
    • Built a bespoke semantic difference tracking system as part of SNOMED International’s framework for quality control of their medical data.

  • Computing Views of OWL Ontologies for the Semantic Web

    • Developed a logic-based, principled approach to creating views of OWL ontologies specified in the description logic ALCHIO.
    • Implemented a prototype of the approach, and compared it with existing tools LETHE and FAME, with results showing better success rates and performance.

  • Adaptive Optimization of Traffic Signal Timing via Deep Reinforcement

    • Proposed a traffic light timing optimization scheme based on deep reinforcement learning, which dynamically adjusts the green light time and phase at an intersection with the goal of minimizing vehicle delay time.
    • Implemented a deep reinforcement learning network (in Python), and reduced waiting time and average queue length in various traffic flow modes by more than 33.4% compared to the traditional timing control.

  • UI-FAME: A High-Performance Forgetting System for~Creating Views of Ontologies

    • Implemented and optimized the UI-FAME system for a non-standard reasoning procedure called forgetting for OWL ontologies specified in the description logic ALC.
    • Designed ontology versioning framework for Babylon ontologies with UI-FAME as back-end, and conducted extensive trials in Babylon set-up with involvement of their software/knowledge engineers.


    Publications

  • UI-FAME: A High-Performance Forgetting System for Creating Views of Ontologies
    Xuan Wu, Wenxing Deng, Chang Lu, Peiqi Wei, Yizheng Zhao and Hao Feng.
    In 29th ACM International Conference on Information and Knowledge Management (CIKM) 2020.
    [PAPER]

  • Computing Views of OWL Ontologies for the Semantic Web
    Jiaqi Li, Xuan Wu, Chang Lu, Wenxing Deng, and Yizheng Zhao. In Proceedings of the Web Conference (WWW) 2021. [PAPER]

  • Adaptive Optimization of Traffic Signal Timing via Deep Reinforcement Learning
    Zibo Ma, Tongchao Cui, Wenxing Deng, Fengyao Jiang, and Liguo Zhang.
    In Journal of Advanced Transportation, vol. 2021, 2021. [PAPER]