CV
Zihan Zhou
Summary
Master's Graduate actively seeking PhD opportunities in Machine Learning and Computer Vision. Research focuses on efficient training methods, 3D reconstruction, federated learning, and continual learning.
Education
- Computer Science2025-05-26Boston UniversityGPA: 3.31
- Computer Science2023-05-18Hobart and William Smith CollegesGPA: 3.15
Work Experience
- Research Assistant2024-08-01 - PresentBoston UniversityAdvisor: Prof. Reza Rawassizadeh. Developing end-to-end camera to 3D human model pipeline using pose estimation, motion reconstruction, and Blender integration. Working on gradient-based early stopping for transformers (GradES paper).
- Co-authored GradES paper achieving 1.57x-7.22x training speedup
- Built complete pipeline from 2D video to 3D character animation
- Implemented LoRA-C fine-tuning for efficient pose estimation
- Research Assistant2025-02-01 - PresentBoston UniversityAdvisor: Prof. Avi Mohan. Investigating federated learning convergence theory through dual-track research combining practical YOLOv8 object detection with controlled theoretical validation.
- Discovered BatchNorm's critical role in FL convergence (18.6% performance impact)
- Built UltraFlwr framework integrating Flower and YOLOv8
- Validated convergence theory across heterogeneity levels (λ from 12 to 2185)
Skills
Deep Learning
- PyTorch
- TensorFlow
- Transformers
- Computer Vision
- Reinforcement Learning
Computer Vision
- 3D Gaussian Splatting
- Pose Estimation
- Object Detection
- YOLOv8
- MediaPipe
Distributed Systems
- Federated Learning
- Flower Framework
- Multi-GPU Training
Research Tools
- Weights & Biases
- Git
- Docker
- Jupyter
- Blender
Publications
- GradES: Significantly Faster Training in Transformers with Gradient-Based Early Stopping2025arXiv preprintNovel gradient-based early stopping method operating at component level within transformers, achieving 1.57x-7.22x training speedup while improving generalization. Co-authors: Qifu Wen, Xi Zeng, Zihan Zhou, Shuaijun Liu, Mehdi Hosseinzadeh, Ningxin Su, Reza Rawassizadeh.
Portfolio
- 3D Scene Understanding and Prediction2025PortfolioResearch on neural 3D scene representations and temporal dynamics modeling. Target: CVPR 2026
- Continual Learning in Sequential Tasks2025PortfolioInvestigating methods to mitigate catastrophic forgetting in sequential learning scenarios. Target: ICLR 2026
- Federated Learning: Theory and Practice2025PortfolioConvergence analysis in federated learning with heterogeneous data distributions. Advisor: Prof. Avi Mohan
- Motion Capture and 3D Character Animation2024PortfolioEnd-to-end pipeline for human motion extraction and 3D character animation. Advisor: Prof. Reza Rawassizadeh
Interests
- Machine Learning EfficiencyGradient-based Early Stopping, Parameter-Efficient Fine-Tuning, Model Optimization
- 3D Vision & World ModelsNeural Representations, Scene Understanding, Temporal Prediction
- Federated LearningConvergence Theory, Heterogeneous Data, Distributed Learning
- Continual LearningCatastrophic Forgetting, Sequential Learning, Multi-Task Learning