cv

Ph.D.

University of Virginia, Charlottesville, USA

Electrical and computer engineering

Advisor: Prof. Cong Shen

M.S.

Texas A&M University, College Station, USA

Electrical engineering

Advisor: Prof. Tie Liu

B.S.

Tsinghua University, Beijing, China

Electrical engineering

Sr. Data Scientist

Juniper Networks.

• Auto channel allocation for AFC-enabled 6G WIFI devices
  • Built the whole pipeline to handle the data flow, auto channel allocation, and error handling for the AFC-enabled 6G WIFI devices.

Research Assistant

University of Virginia, Charlottesville, USA

• Best Arm Identification for Prompt Learning under a Limited Budget
  • Built the connection between the prompt learning problem for large language models and the best arm identification problem in multi-armed bandit.
  • Proposed a general scheme to solve the prompt learning problem with a limited budget.
  • Provided two context-based algorithms for prompt learning with a limited budget using clustering and function approxiamtion.
  • Tested the performance of the proposed algorithms over diversed tasks and different LLMs (both black box and white box).
• Reinforcement learning for wireless system optimization
  • Designed a new structure of Multi-agent reinforcement learning for wireless user scheduling.
  • Comprehensively tested the performance, transferability and robustness of MARL and centralized RL for wireless user scheduling.
  • Tested the performance of SOTA offline RL algorithms for wireless user-scheduling problem.
  • Designed a dataset mixture scheme to improve the training performance using sub-optimal quality datsets.
  • Designed an ensemble scheme to improve the performance of dataset mixture for offline RL.
  • Provided an insight on when the mixture of datasets will improve the performance of offline RL.
  • Extended the offline RL framework to more general network optimization problems like network-slicing.
• Server-client collaboration in Federated learning
  • Designed a new algorithm to make Server and Clients in Federated learning to work collaboratively.
  • Proved the improved convergence speed of our algorithm with designed learning rate.
  • Tested the performance of our algorithm with real-world dataset and edge-devices.

Research Scientist Intern

Intel Corporation, Portland, USA

• NS3-based communication simulator for OpenFL
  • Co-designed the structure of NS3-based communication simulator for OpenFL.
  • Enabled the Carla simulator to communicate with the NS3-based communication simulator under federated learning setting.
  • Tested the performance of the SOTA FL algorithms under new communication module.
• Reinforcement learning for Network Slicing resource allocation
  • Enabled RL training scheme for network slicing resource allocation using Intel's in-house networkgym simulator.
  • Tested the performance of the SOTA algorithms under different network settings.
  • Boosted the performance of rule-based method for ~30% using RL agents.

Research Scientist Intern

Xsense.ai, San Diego, USA

• Deep learning based behavior prediction
  • Designed a VectorNet+Residual network structure for behavior prediction. Achieved comparable performance with less parameters.
  • Investigated the performance of different feature encoding mechanisms for behavior prediction.
  • Re-implemented the SOTA behavior prediction algorithm TNT and VectorNet using Pytorch on a in-house dataset and delivered live demo.

Research Scientist Intern

Kneron, San Diego, USA

• Adaptive deep neural network quantization toolbox
  • Surveyed and summarized the most recent and popular neural network quantization and compression methods and schemes.
  • Developed an adaptive tool box for deep neural network quantization that can automatically select the best quantization accuracy for different layers in pre-trained deep learning models.
  • Quantized popular deep learning models with the toolbox that achieved less than 2% of accuracy loss with a roughly 50% computational cost reduction.

Fellow

Insight Data Science, San Francisco, USA

• Monitoring stack for machine learning system
  • Worked with gaming analytics company, Mayhem, to build a robust monitoring stack.
  • Launched whole system onto Mayhem's real server system
  • Deployed a clone of Mayhem's server and the monitoring stack on to AWS with Terraform.
  • Collected enhanced monitoring metrics from Mayhem's RDS Mysql cluster using Prometheus.
  • Designed a dashboard with Grafana that can accurately reflect resource allocation and usage of a RDS database and the error and slow query log flow.
  • Created alarms for potential spike traffic and resource shortage for the database.

Software Engineer Intern

Xsense.ai, San Diego, USA

• Deep reinforcement learning based ramp merging control
  • Developed a new feature for real traffic simulator that enabled the machine learning models to control the ego vehicle directly with C++ or Python scripts, which is used by the research team for the company.
  • Created an API that allowed the ROS simulator to sample and log data with a particular frequency using python. This API reduced more than 80% of a potential memory waste caused by the ROS logging system.
  • Designed a new reward function for the reinforcement learning ramp merging that treated safety as a part of reward instead of a hardcoded boundary while still maintain a hard boundary like behavior during training.
  • Built a DDPG agent based on Tensorflow using python. The agent was able to achieve a 0.24% collision rate with our reward function, comparable to the state-of-art method S-T optimizer.