cv
Basics
| Name | Can Jin |
| Label | Ph.D Student@Rutgers University |
| can.jin@rutgers.edu | |
| Phone | (908) 202-5264 |
| Url | https://jincan333.github.io/ |
| Summary | I'm a Computer Science Ph.D. student, starting in Fall 2024 at Rutgers University, New Brunswick, under the guidance of Professor Dimitris N. Metaxas. My research interests include Large Foundation Models, Efficient AI, and AI Agents. |
Work
-
2025.06 - 2025.09 San Jose / Remote, CA
Research Scientist Intern
Adobe Incorporated
Large Foundation Models Pretraining
- Identify critical limitations in existing Top-p routing for Mixture-of-Experts (MoE), specifically regarding uncontrolled sparsity and sensitivity to hyperparameter selection.
- Propose DTop-p MoE, a dynamic routing scheme that learns the probability threshold via a PI controller, introducing dynamic routing normalization to ensure expert diversity.
- Demonstrate through comprehensive NLP and CV experiments that DTop-p consistently outperforms Top-k and Top-p MoE across varying expert granularities, expert capacities, model sizes, and dataset sizes.
-
2023.01 - 2023.12 Remote
Research Intern
Massachusetts Institute of Technology (MIT)
Efficient AI
- Conduct a pilot study on post-pruning visual prompts, revealing its inefficacy in improving the performance of well-fine-tuned sparse vision models.
- Develop VPNs, a novel data-model co-design paradigm that simultaneously optimizes weight masks and visual prompts for efficient sparse vision models.
- Achieve state-of-the-art efficiency-performance trade-offs across diverse datasets, architectures, and pruning regimes (LTH, OMP, GraSP, SNIP, SynFlow, HYDRA, and BiP), validated by extensive empirical results.
-
2020.07 - 2023.08 Shanghai, China
Machine Learning Engineer
Meituan Dianping Corporation
Sales Forecasting and Time Forecasting
- Sales Forecasting: Developed XGBoost and DNN models with quantile prediction to forecast SKU sales, increasing prediction accuracy by 4% (92% → 96%) and optimizing inventory management.
- Time Forecasting: Constructed Attention-based DNNs to estimate warehouse task duration, improving accuracy by 9% (74% → 83%) and significantly reducing labor expenditures and heightening workforce efficiency.
Education
-
2024.08 - Present New Brunswick, NJ
Ph.D. Candidate
Rutgers University
Computer Science
- Pretraining/Posttraining/Inference of Large Foundation Models
- Efficient AI
- AI Agents
-
2018.09 - 2020.07 Hefei, China
Master of Science
University of Science and Technology of China (USTC)
Mathematics
- Stochastic Partial Differential Equations (SPDEs)
-
2014.09 - 2018.07 Hefei, China
Bachelor of Science
University of Science and Technology of China (USTC)
Mathematics
- Probability Theory
- Random Processes
- Dynamical Systems
Awards
- 2025
- 2020
- 2020
- 2015
Skills
| Languages & Tools | |
| Python | |
| PyTorch | |
| C/C++ | |
| SQL | |
| Git | |
| Slurm | |
| Docker | |
| Cursor | |
| VS Code | |
| RStudio | |
| MATLAB |
Languages
| English | |
| Fluent |
| Chinese | |
| Native speaker |
Interests
| Large Foundation Models | |
| Pretraining | |
| Posttraining | |
| Inference | |
| Generalization | |
| Reasoning |
| Efficient AI | |
| Model Compression | |
| Distillation | |
| Pruning | |
| Prompting |
| AI Agents | |
| Multi-agent Systems | |
| Reinforcement Learning |
Volunteer
-
2024.09 - Present Teaching Assistant
Rutgers University
CS534: Computer Vision (Spring 2025), CS210: Data Management for Data Science (Fall 2024), CS211: Computer Architecture (Fall 2025)
-
Reviewer
Journal Reviewer
Alexandria Engineering Journal, Information Fusion, Pattern Recognition, Signal Processing
-