CV
You can also find my CV here: Jinghao Zheng’s Curriculum Vitae.
📖 Education
2025.08 - now, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland M.S. in Computer Science.
2021.09 - 2025.06, Shanghai Jiao Tong University (SJTU), Shanghai, China
B.E. in Automation(Computer Science and Engineering).
- Major GPA:3.82/4.3
- Centesimal grade average:89.23/100
- Core Courses: Calculus ΙΙ (98), Probability and Statistics (99), Linear Algebra (92), Discrete Mathematics (93), Data Structure (90), Pattern Recognition (96), Principles of Automatic Control (94), Robotics (93)
📝 Publication
Do We Need All the Synthetic Data? Targeted Image Augmentation via Diffusion Models, Dang Nguyen†, Jiping Li†, Jinghao Zheng†, and Baharan Mirzasoleiman. ICLR 2026 (Accepted as a poster)
Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement, Z. Huang, X. Cheng, J. Zheng, H. Wang, Z. He, T. Li, and X. Huang. NeurIPS 2024 (Accepted as a spotlight)
🔬 Research Experience
- Feb. 2026 – Present: Research Assistant
- Project: Prompt-Faithful Video Generation with Dynamic Sink Adjustment
- Investigated sink-based forcing methods in video generation and found that sink mechanisms used to reduce drift can degrade prompt following for salient foreground objects.
- Designed and implemented a dynamic sink adjustment method that predicts the spatial locations of key foreground objects from prompts and adapts sink placement through perturbations to the corresponding spatial RoPE vectors.
- Improved prompt following by better aligning foreground object generation and spatial layout with textual prompts.
- Advisor: Alexandre Alahi, Associate Professor, School of Engineering, EPFL
- Project: Prompt-Faithful Video Generation with Dynamic Sink Adjustment
- Nov. 2024 – June. 2025: Research Assistant
- Project: Efficient Continual Learning for LLMs: A Parameter Sorting Approach to Mitigate Catastrophic Forgetting
- Proposed a dynamic block-wise parameter sorting method that identifies and protects task-critical parameters during fine-tuning, effectively mitigating catastrophic forgetting in a continual learning setting for LLMs.
- Dynamically updated a subset of parameters to preserve performance on previous tasks while improving training efficiency compared to existing methods.
- Advisor: Xiaolin Huang, Professor, Vice Dean, Department of Automation, SJTU
- Project: Efficient Continual Learning for LLMs: A Parameter Sorting Approach to Mitigate Catastrophic Forgetting
- July 2024 – May 2025: Research Assistant
- Project: Targeted Synthetic Image Augmentation via Diffusion Models
- Contributed to building a selective synthetic data augmentation framework based on GLIDE text-to-image model, which identifies “slow-learnable” examples early in training and augments only this targeted subset.
- Designed a faithful image generation pipeline that initializes the diffusion process from real samples to produce synthetic images that preserve semantic features while diversifying noise patterns.
- Conducted experiments across multiple architectures (ResNet, ViT, ConvNeXt, Swin) and datasets (CIFAR-10/100, TinyImageNet); augmenting only 30%–40% of training data achieves up to 2.8% accuracy gain, while reducing training overhead by 70% vs. 2× full-dataset diffusion augmentation.
- Advisor: Baharan Mirzasoleiman, Assistant Professor, Computer Science Department, UCLA
- Project: Targeted Synthetic Image Augmentation via Diffusion Models
- Mar 2024 – June 2024: Research Assistant
- Project: Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement
- Proposed replacing Euclidean distance with KL divergence on the remaining set as a retention constraint, better respecting the geometry of the output probability space and improving retained model performance during unlearning.
- Conducted experiments and hyperparameter tuning on image classification and generation tasks, with systematic comparison against multiple unlearning baselines.
- Achieved an average 1.8% disparity reduction on CIFAR-10 random-subset forgetting and an average FID improvement of 80 on ImageNet class-forgetting.
- Advisor: Xiaolin Huang, Professor & Vice Dean, Department of Automation, SJTU
- Project: Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement
- Feb. 2024 – June 2024: Sole Researcher
- Project: Polyp Detection and Segmentation Augmented by Diffusion Model
- Implemented yolov10 and ResUnet++ as baselines to finish object detection and segmentation on medical images.
- Proposed using Diffusion-based generative models to generate synthetic data for data augmentation, which improved the mAP0.5@0.95 in the object detection by 1% and the IoU in the segmentation by 5%.
- Advisor: Manhua Liu, Professor, Artificial Intelligence Research Institute, SJTU
- Project: Polyp Detection and Segmentation Augmented by Diffusion Model
- Oct. 2023 – Mar. 2024: Group Member
- Project: Design of distributed collaborative positioning system
- Developed motor control code on the STM32 board to precisely manage the yaw and pitch of the camera platform, ensuring accurate angle adjustments.
- Contributed to the mechanical design of the camera head and designed circuit boards to interface the STM32 board with multiple cameras.
- Advisor: Jianping He, Associate Professor, Department of Automation, SJTU
- Project: Design of distributed collaborative positioning system
- Mar. 2023 – Feb. 2024: Project Leader
- Project: Implementation and comparison of gas tracing algorithms for dual robots in confined space
- Proposed a bionics-based gas tracing algorithm for dual robots in a confined space and conducted experiments to simulate and validate our algorithm, which improved the success rate by 1.5% and the search efficiency by 9%.
- Developed control code for Raspberry Pi to ensure precise movement and implemented ROS2 communication protocols for real-time data exchange between the robots and the main computer.
- Advisor: Liufang Wang, Senior Engineer, Student Innovation Center, SJTU
- Project: Implementation and comparison of gas tracing algorithms for dual robots in confined space
🎖 Honors & Awards
- Outstanding Graduate of SJTU, 2025
- 3rd Prize of TI Cup National Undergraduate Electronic Design Contest Shanghai area, 2023
- 3rd Prize of Academic Scholarship of Shanghai Jiao Tong University, 2022 & 2023 & 2024
- 2nd Prize of Chinese Physics Olympiad (Zhejiang Area), 2019
🛠️ Skills
- Programming Languages: Python, C/C++, ROS2, Matlab, Markdown, LaTex
- Languages: Chinese (Native), English (Proficient)
- Leadership Experience: Head of Sports Department, School of Electronic Information and Electrical Engineering (SEIEE) Student Union, Shanghai Jiao Tong University