KOSEN 한인과학기술자네트워크

Research

• In-memory Engine
  • RTI + Graph
    • Multi-model representation and fusion analytics
    • Unified transaction management on divergent data model
    • Dynamic graph data structure for real-time analytics and fast transaction support
• Intelligence Data Engineering
  • Model & Data Lifecycle Management
    • Deep learning, the mainstream artificial intelligence research topic is using a lot of computing resources like GPU, FPGA, and ASIC. We need to handle this resources efficiently in the cycle process including continuous using, getting feedback, doing update the models. The goals of this research are 1) to define the lifecycle of deep learning model and data, 2) to find some engineering problems in the lifecycle management and 3) to implement the management system.
  • RTI + Deep learning framework
    • The goal of this research is to tightly integrate deep learning libraries into our data processing platform. Data collection and preprocessing is performed in this single framework without overhead. This approach can contribute to optimized performance by removing unnecessary data movement and alleviate complexity of big data analysis system.
• BigData Application
  • Urban Problem Solving
    • Air pollution forecasting
      • Air pollution is an urgent problem in South Korea. It is caused by many factors such as transportation vehicles, industrial plants, China impacts, and so on.
      • Datasets : collect hourly air pollution data from monitoring stations around major cities.
      • Use Recurrent Neural Network (RNN) Encoder-Decoder model with an attention layer. Combine both Air pollution values and Weather data in the model to predict ahead 8 → 24h Air pollution value.
  • Planning
    • Planning is the process of thinking about and organizing the activities required to achieve a designed goal. When executing the planning, we move from the initial state to many different states to reach the goal state. Our objective of research is to find the best path to move from initial state to the goal state. By collecting automatically and in real-time actual execution logs, combining with analysing historical execution data, we could apply machine learning and/or deep learning to predict the best path for planning execution.