About me

I'm an Applied Scientist at Microsoft Turing team.

Research Interests

  • LLM icon

    Large Language Model

    • Parameter Efficient Finetune
    • Question Answering
    • Reinforcement Learning
  • NLP icon

    Natural Language Processing

    • Multilingual Machine Translation
    • Efficient Token Representation
    • Privacy Preservation
  • FL icon

    Federated Learning

    • Private Federated Learning
  • RL icon

    Reinforcement Learning

    • Multi-agent Reinforcement Learning

My skills

Programming languages
Python
Java
C
MATLAB
CSS
HTML
Deep Learning
PyTorch
TensorFlow
Numpy
Pandas
Hugging Face
Tools
Git
Linux
LaTeX
AWS

Education

  1. Stevens Institute of Technology
    Hoboken, NJ, USA
    August, 2019 — May 2024

    • Ph.D. in Computer Science
    • Reseach areas: Large Language Model, Natural Language Processing, Federated Learning

  2. Xidian University
    Xi'an, Shaanxi, China
    September, 2016 — June 2019

    • M.S. in Signal and Information Processing
    • Reseach areas: Bayesian Models, Deep learning

  3. Xidian University
    Xi'an, Shaanxi, China
    August, 2012 — June 2016

    • B.S. in Electronic Engineering (Education reform class)

Work

  1. Microsoft
    Redmond, WA, USA
    July 2024 — Now
    Applied Scientist
  2. Adobe
    San Jose, CA, USA
    May 2023 — December 2023
    Research Scientist/Engineer Intern
  3. Stevens Institute of Technology
    Hoboken, NJ, USA
    Teaching Assistant

    • Course: CS560 Statistical Machine Learning (Graduate)
      January 2024 — May 2024
    • Course: CS541 Artificial intelligence (Graduate)
      September 2023 - December 2023
    • Course: CS583 Deep Learning (Graduate)
      January 2023 - May 2023
    • Course: CS583 Deep Learning (Graduate)
      September 2022 - December 2022
    • Course: CS284D Data Structure (Undergraduate)
      January 2022 — May 2022

Projects

  1. Adobe
    San Jose, CA, USA
    May 2023 — December 2023
    Research Scientist/Engineer Intern

    • Project: Finetune and evaluate Large Language Models (LLMs) on domain-specific Question Answering (QA) data.
    • Main skills: Python/PyTorch/LLMs/Git/Question Answering/Pandas/Bash
    • Experiments: Fine-tune Falcon 7B and Falcon 40B models for QA; Employ robust evaluation methodologies, utilizing industry-standard metrics such as ROUGE and METEOR for automatic assessment. Develop an innovative LLM-based evaluation metric, assessing both semantic similarity and correctness with the LLMs.
    • Results: Demonstrate an approximate enhancement of 0.1-0.2 across key automatic evaluation metrics including ROUGE-1, ROUGE-L, and METEOR. Obtain an increase of ~1.0 (on a scale of 1-5) in terms of both similarity and correctness, as measured by the GPT-4 evaluator.

  2. Stevens Institute of Technology
    Hoboken, NJ, USA
    September, 2022 — December 2023
    Research Assistant

    • Project: Private NLP Model in Federated Learning
    • Task: Protect users’ privacy against attacks based on embedding gradients with bytes.
    • Experiments: NLP tasks including machine translation, sentiment analysis, and language modeling tasks.
    • Results: Increased approximately 1.0 BLEU point on translation and up to 1.3 accuracy on sentiment analysis over the baseline model. Defended against attacks based on subword inference from the gradients while maintaining model performance and efficiency.

  3. Stevens Institute of Technology
    Hoboken, NJ, USA
    December 2021 — August 2022
    Research Assistant

    • Project: Byte-based Multilingual Machine Translation
    • Task: Improve multilingual machine translation based on byte tokenization
    • Experiments: NLP tasks including machine translation, sentiment analysis, and language modeling tasks.
    • Results: Increased up to 18.5 BLEU points of translation on low-resource and endangered languages. Enhanced the generalizability and robustness of the byte-based model with our proposed random byte encoding with ensemble prediction.

  4. Stevens Institute of Technology
    Hoboken, NJ, USA
    September 2019 — November 2021
    Research Assistant

    • Project: Privacy & Security in Federated Learning
    • Task: Propose a defense method for secure collaborative learning with matrix sketching.
    • Experiments: Classification of MNIST and CIFAR-10 under federated learning settings and input recovery with and without our defense.
    • Results: Proved the effectiveness of our defense theoretically and experimentally. Protected user privacy without compromising model performance. The per-communication round complexity is reduced to 0.5x. The L-2 Norm of gradient matching loss with our defense method increases from 0 to 25 – 100, making the attacks much more difficult.