Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains
Edward Zhang et al. arXiv:2608.13729, 2026 · First author
About
I’m a third-year Ph.D. student in Computer and Information Science at the University of Pennsylvania’s GRASP Laboratory, advised by Eric Eaton and Dan Hashimoto. I am an NSF Graduate Research Fellow. My research sits at the intersection of computer vision, robotics, multimodal foundation models, and generative learning, with a particular interest in specialized, data-scarce, and safety-critical environments.
Computer Vision & Robotic Perception
Many machine-learning methods assume that representative data can be collected when needed. In safety-critical settings, important examples may instead be intrinsically scarce: severe injuries, genuinely rare events, or events that have not happened yet. I study the features that make these sparse examples distinct and how to expand those regions of a distribution. This also means asking what foundation models do not already represent: when abundant examples exist, simple discriminative models can often learn the concept. My current work explores more targeted retrieval, generation, and data expansion without merely memorizing scarce examples.
Publications
Selected work on robotic triage, specialized visual recognition, and embodied perception.
A Multi-Robot Platform for Robotic Triage Combining Onboard Sensing and Foundation Models
Jason Hughes, Marcel Hussing*, Edward Zhang*, et al. arXiv:2512.08754, 2025 · *Equal contribution
Benchmarking Visual Language Models for Automated Trauma Detection in Triage Scenarios
Edward Zhang, Marcel Hussing, Jason Hughes, et al. AAAI-25 Workshop on Artificial Intelligence in Healthcare, 2025
Ego-Exo4D
CVPR 2024
open-body-fit
MIG 2022 poster
Research & Projects
Current work in robotic perception and data-scarce vision, followed by earlier work in embodied data collection, neural rendering, graphics, and game development.
- All
- Research
- Graphics Programming
- Game Development
DARPA Triage Challenge / Robotic Perception
Developing perception for autonomous casualty assessment and triage: body-part perception, injury classification, trauma understanding, and alertness-related perception. I work with conventional vision models and VLMs; deploy ROS/ROS2 and Dockerized components on real robotic platforms; work with field data, sensors, and hardware constraints; coordinate technical resources; and mentor approximately four master’s students.
Synthetic Data for Specialized Data-Scarce Vision
First-author work benchmarking generative and conventional augmentation for specialized trauma recognition on held-out real images. I examined distribution shift, memorization/collapse, and overly canonical synthetic examples; the evaluated generative methods did not consistently exceed a strong non-generative baseline. Ongoing work asks how to target the features that define sparsity without memorizing scarce examples.
