Experience

Keywords

  • Computer vision and big data
  • C++ modern programming
  • Full-stack application

Topic: object re-identification and tracking

Back in 2019 when I was in Munich, I joined the deep computer vision lab to work with kalman filter variations. I expanded my object re-identification and tracking experience starting in 2021 at Columbia. I first got started with VAE-(W)GAN to generate synthetic images and label smoothing with outliers(CVPR 2017) and poly-loss(ICLR 2022). In early 2022, I joined Kostic lab for a while to investigate BoTs in self-supervised re-identifications with model zoos, both in pedestrian and vehicles, and at the same time, to build a machine learning pipeline with data auto-fetch by tuned YoloV4. I applied my experience in NVIDIA metropolis team in 2022 summer, and integrated the algorithm into a streaming data pipeline.

Topic: High-performance software development

I have experience in high-performance software development. I started with CUDA programming to accelerate KMeans with a single GPU. I got in touch with model quantization at Columbia and also participated in Megvii’s mini-workshop and gained first place. In late 2022, I developed faster multi-GPU training on both CNN and transformer models. I also tried to adapt to C++20’s new features, multi-threading, and concurrency in my recent full-stack development based on the course taught by Prof. Stroustrup.

Projects cover a wide range of topics, where you can get access to the details through GitHub as well as my CV.

Projects

  • Strong Reinforcement Baseline on Atari Skiing w/. Immitation Learning
  • Real-time Self-supervised Re-identification Algorithm and BoTs
  • Real-time Single Camera Re-identification w/. Deep Convolutional GAN and BoT
  • New York Traffic Heatmap w/. Google API, Tomtom Traffic and Spark Streaming
  • Transformer Full Quantization for Text Classification Task
  • Full Stack Application Demo w/. Modern C++ and Sqlite
  • …

Selected Resources

Open-Sourced Slides/Reviews

Presentations