Research Experience

• 08/2024-present: Research Assistant
  • Design a Novel compression algorithm for improving distributed GNN training process
    • Indiana University Bloomington
    • Duties includes: Worked on developing a communication-efficient compression algorithm to accelerate distributed Graph Neural Network (GNN) training. Focused on reducing communication overhead during feature and gradient exchange across nodes, while maintaining model accuracy. Contributed to algorithm design, performance analysis, and validation on large-scale GNN benchmarks.
  • Design CXL-based system for cost-efficient multi-LoRA serving
    • University of California, San Diego
    • Helped profile GPU performance under multi-LoRA inference workloads and identified key bottlenecks to support system-level optimization. Created visualizations of memory and compute utilization patterns to assist in the design of a CXL-enabled architecture for large-scale, cost-efficient LLM inference.
  • Analog Simulation for a Novel Machine Learning Design
    • University of Rochester
    • Duties includes: Leveraging the natural evolution of dynamical systems toward their optimal steady energy state- a property that can be harnessed to solve optimization problems. We designed a specialized analog circuit to accelerate both training and inference processes. To validate our initial design, we used capacitors and resistors to represent nodes and weights, connecting them via a dedicated circuit. In simulation, the system evolved to a final state that served as the inference result in the prototype.
  • Analog Simulation for a Novel Machine Learning Design
    • Indiana University Bloomington
    • Duties includes: Leveraging the natural evolution of dynamical systems toward their optimal steady energy state- a property that can be harnessed to solve optimization problems. We designed a specialized analog circuit to accelerate both training and inference processes. To validate our initial design, we used capacitors and resistors to represent nodes and weights, connecting them via a dedicated circuit. In simulation, the system evolved to a final state that served as the inference result in the prototype.
  • Implementation onto FPGA for Scaling Data Compression Algorithm
    • Indiana University Bloomington
    • Duties includes: SZ compression algorithm is a lossy compression algorithm used for large data storage and communication. We designed an innovative algorithm in quantized stage to improve the parallelism for SZ3.0 compression algorithm, where a special architecture is designed to implement this improved algorithm on FPGA. As a result, a prototype as this initial feasibility architecture is verified.
• 07/2021-05/2023: Research Assistant
  • Study on the MEMS Fast Steering Mirror Control System
    • University of Chinese Academy of Sciences
    • Duties includes: Designed a unique charge amplification circuit to drive the MEMS FSM based on its special electrical structures, ensuring the matching of electric parameters. This solved the attenuation problem of piezoelectric cantilever beams elongation during discharging. Conducted thorough research on other MEMS mirror testing systems, leading to the establishment of a proposed system to characterize frequency and open-loop bandwidth of the MEMS FSM.
  • Research on the Control System of the MEMS Fast Steering Mirror
    • University of Chinese Academy of Sciences
    • Duties includes: Proposed an improved double-step algorithm to enhance the quasi-static performance of the MEMS FSM, reducing the settling time from 398 ms to 0.4 ms, effectively countering satellite platform vibrations below 500 Hz. Improved tracking precision, reducing repeat positioning error from 200 μrad to 30 μrad, surpassing other MEMS FSMs.
  • Research on the Diamond Magnetic Sensor Miniaturization
    • University of Chinese Academy of Sciences
    • Duties includes: Designed external circuitry to integrate the MEMS magnetic field sensor and connect the electric pins. Developed and optimized a digital filter to address excessive initial signal noise, leading to effective noise suppression and an overall improvement in signal quality.
  • Research on the MEMS Laser Radar Imaging System with Lissajous Scanning
    • University of Chinese Academy of Sciences
    • Designed a digital PID controller to stabilize the operation of the MEMS mirror at the desired frequency for autonomous driving imaging systems. Conducted imaging experiments at different resolutions using the system I constructed.