Hey, I'm Xiaocen

A rising senior student in University of Michigan major in Computer Science and Financial Math.

profile

Who I am

I'm currently pursuing Bachelor of Computer Science and Financial Math in University of Michigan, Ann Arbor, expected to graduate in May 2025. Before that, I went to UC, San Diego and transfered to U-M in 2022. They have totally different weather, but I love both of them!

My track focuses on backend software(looking for SDE intern for 2024 summer!), machine learning intern. My research interest focuses on machine learning theory and statistical learning. I'm always curious about knowledge and enjoying exploring the theoretical explanation for my observation.

I'm Chinese,and speaks Mandarin as my native language(my Chinese name is 刘笑岑). When I’m not working or studying, I focus on my hobbies which are: training in gym, playing piano, go hiking and hanging out with my friends!

Experiences

University of MichiganResearch Assistant
University of Michigan
Under Professor Wei Hu
  • Explored linear mode connectivity in Transformer models with pytorch, assess relation between stability and generalization.
  • By interpolating the weights of two copies of networks trained from initialization and from pretrained k epochs network, we calculate the supremum difference as error barrier to determine the instability.
Tools: PyTorch, Network Interpolation Techniques
Feb 2024 - Present | Ann Arbor, MI
CPIC Fund Management Co LtdSoftware Engineer Intern
CPIC Fund Management Co Ltd
A leading funding company in China
  • Develop a NLP project to classify textual news titles(sentimental analysis) for stock prediction, improve accuracy by 4% overall.
  • Leveraged ChatGPT API to assist improving accuracy, experiment training with Transformer, ResNet-10, and LSTM, reach test accuracy 93%, perplexity 62.3.
  • Implemented Importance Sampling - SGD for underrepresented news types, improves underrepresented accuracy by 30%.
  • Process 100G data by utilizing PostgreSQL, using Dask for parallel computing. Integrated Docker for CI/CD pipelines, integrated this feature with REST API. Analyze and cleanse data using numpy, panda, and tensorflow, present results using Powerpoints.
Tools: NLP, ChatGPT API, Transformer, ResNet-10, LSTM, Importance Sampling - SGD, PostgreSQL, Dask, Docker, REST API, Numpy, Pandas, TensorFlow
June 2023 - August 2023 | Shanghai, China
University of MichiganResearch Assistant Intern
University of Michigan
Under Professor Honglak Lee
  • Engaged in reinforcement learning project by modifying DIAYN algorithm with Jax and flax framework, to analyze how two objects interact with each other, an off-policy, unsupervised learning, efficiency with parallel programming on GPU.
  • Evaluated transfer learning methods using the distributional generalization as the algorithm design principle.
  • Experiments in varying environment and task scenarios in OpenAI Gym, improved time to convergence by 20%.
Tools: Jax, Flax, DIAYN, GPU Parallel Programming, OpenAI Gym
April - July 2023 | Ann Arbor, MI

Skills

Languages

PythonPython
C++C++
CC
JavaJava
PostgreSqlPostgreSql
JavaScriptJavaScript

Libraries and Frameworks

React.jsReact.js
TensorFlowTensorFlow
NumpyNumpy
PandasPandas
PyTorchPyTorch
JaxJax
FlaskFlask
bootstrapbootstrap

Technologies

AWSAWS
Apach HadoopApach Hadoop
DockerDocker

Projects

Ann Leasing

  • Led a team of 3 and developed a full stack web application, for subletting aparments, with features like search, listings, user authentication, emphasizing responsiveness and security, hosted on AWS, using BeautifulSoup to enrich the listings.
  • Utilized React.js and Tailwind CSS for the frontend, Node.js, Flask, and REST API for the backend, and PostgreSQL for data management, managing user authentication and authorization using Firebase.
Descriptive Alt Text

Handmade Search Engine

  • Engineered a full stack scalable search engine comparable to Google/Bing, integrating text analysis (tf-idf), link analysis (PageRank), and MapReduce with Hadoop framework for data processing.
  • Implemented a segmented inverted index and a REST API Index server for efficient search result retrieval in JSON format.
  • Designed a user-friendly Search server interface with CSS, providing search capabilities similar to major search engines.
Descriptive Alt Text

Yoga Pose Correctness

  • Designed and implemented a full stack (hardware+software) computer vision product to recognize/evaluate a user's yoga gesture from Pi camera stream, with image/video being processed by Tensorflow and OpenCV on raspberry pi 4.
  • Classify them into four basic poses using MediaPipe body tracking, resulting in program correctness of around 85%, using QT framework to improve user-end usability.
Descriptive Alt Text

Get in touch

Please feel free to reach out by email at xiaocen@umich.edu!