About Me
I'm Zhen Tong, a first-year graduate student at the Carnegie Mellon University, MSIN. My passion for machine learning has driven me to gain substantial experience in diverse areas such as game AI, image generative models, and various hands-on projects. Throughout my career, I have had the opportunity to work with a wide range of machine learning models and techniques, which has allowed me to build a solid foundation in Python and PyTorch. I have experience in implementing and optimizing algorithms for tasks including reinforcement learning for game AI agents, developing image generative models, and designing backend services for AI applications.
Work Experience
- Designed efficient game state and action representations for a series of poker games, enabling effective reinforcement learning.
- Developed a distributed ETL (Extract, Transform, Load) pipeline using Ray to prepare large-scale game data in parquet format for model training.
- Implemented a deep imitation learning algorithm that achieved an 80% win rate against human experts.
- Deployed the trained agent and conducted extensive client-side concurrency performance tests.
- Proposed a dataset spanning 3-4 kilometers of road segments with incomplete lane lines to address challenges in cost-effective, real-world lane-level HD map generation.
- Designed a Double-Stage Fusion model that first uses a VQGAN to reconstruct input images, then trains a Transformer to learn the latent conditional distribution for generation.
- Evaluated the model’s performance on our EcoMap dataset, demonstrating significant improvements in lane-level HD map generation compared to existing approaches.
- Presented the novel benchmark and baseline model at NeurIPS 2024, contributing to the advancement of the field.
- Selected useful data in the big data system using MySQL, organized all relevant data pertaining to the problem, and successfully aligned the useful data into a new real-time dataset.
- Trained the LSTM-RNN model using PyTorch, predicting future system states based on the collected data. The model demonstrated a performance accuracy of 92%.
- Devised and deployed the infrastructure in the big data system through Hadoop, then created efficient algorithms to run the prediction model, enabling scalability to handle large amounts of data.
- Set up a WebSocket back-end using Java with Maven and handled visualization requests from the front-end, resulting in the rendering of data in the user interface.