PhD Candidate in Biostatistics at University of North Carolina at Chapel Hill
Embedded Files
My name is Kevin (Chen) Wang, and I am a fourth-year PhD candidate in Biostatistics at the University of North Carolina at Chapel Hill, where I am advised by Professor Didong Li. My research lies at the intersection of nonparametric statistics and machine learning theory, with a particular focus on theoretical guarantees for recursive learning systems.
Research Interests
Non-parametric Regression
Machine Learning Theory
Spline Smoothing
Reproducing Kernel Hilbert Space methods
Recursive Training
Manifold Geometry in Stat/ML
Gaussian Processes
Mixed Effects Models
Li Lab
Li Lab
Highlighted Papers
Deep Generative Models: Complexity, Dimensionality, and Approximation
We establish theoretical guarantees showing that deep generative networks can approximate distributions supported on low-dimensional manifolds even when the latent dimension is smaller than the intrinsic manifold dimension. The results reveal a fundamental trade-off between approximation accuracy, latent dimensionality, and network complexity, challenging conventional interpretations of the manifold hypothesis. Published in the Journal of Machine Learning Research (2025).
Can Generative AI Survive Data Contamination?
This work develops a general theoretical framework for recursive training, where AI models learn from mixtures of real and synthetic data. We show that model collapse is not inevitable as long as incoming data samples include real data: under broad conditions, recursively trained models converge to the true distribution, with convergence rates determined by both model quality and the availability of new real data. Experiments on images and text demonstrate stable learning even in heavily contaminated environments. preprint
All Papers
Deep Generative Models: Complexity, Dimensionality, and Approximation
Deep Generative Models: Complexity, Dimensionality, and Approximation
Kevin Wang, Hongqian Niu, Yixin Wang, Didong Li
Journal of Machine Learning Research, 2025.
This work studies the relationship between latent dimension, manifold structure, and approximation power in deep generative models. We show that generative networks can represent distributions on manifolds using latent spaces of arbitrary dimension, while quantifying the resulting complexity–accuracy trade-off.
Can Generative Artificial Intelligence Survive Data Contamination? Theoretical Guarantees under Contaminated Recursive Training
Can Generative Artificial Intelligence Survive Data Contamination? Theoretical Guarantees under Contaminated Recursive Training
Kevin Wang, Hongqian Niu, Didong Li
arXiv, 2026.
This work develops a general theoretical framework for recursive training, where AI models learn from mixtures of real and synthetic data. We show that model collapse is not inevitable and establish conditions under which recursively trained models remain stable and converge to the true data-generating distribution.
Otolaryngology Residency Application Trends and Match Outcomes at a Large Academic Medical Center
Otolaryngology Residency Application Trends and Match Outcomes at a Large Academic Medical Center
Ezer H. Benaim, Lauren M. Cook, Aurelia Monk, Cameron P. Worden, John B. Henrich, Kevin C. Wang, Didong Li, Rupali N. Shah, Adam J. Kimple, Robert A. Buckmire, et al.
ORL, 2025.
This study examines residency application and match outcomes in otolaryngology, providing insights into applicant behavior, institutional trends, and factors associated with successful residency matching.
Identifying the Surgical to Nonsurgical Care Transition for Trauma Patients: Time to Redesign Trauma Workflow
Identifying the Surgical to Nonsurgical Care Transition for Trauma Patients: Time to Redesign Trauma Workflow
Ursula Adams, William Yu Luo, Kevin Chen Wang, Didong Li, Pascal Osita Udekwu, Anthony Charles
The American Surgeon, 2026.
This work investigates the transition of trauma patients from surgical to nonsurgical management and identifies opportunities to improve workflow design, care coordination, and resource utilization in trauma systems.
SARS-CoV-2 Neutralizing Antibody Response After Three Doses of mRNA1273 Vaccine and COVID-19 in Hemodialysis Patients
SARS-CoV-2 Neutralizing Antibody Response After Three Doses of mRNA1273 Vaccine and COVID-19 in Hemodialysis Patients
Xiaoling Wang, Maggie Han, Lemuel Rivera Fuentes, Ohnmar Thwin, Nadja Grobe, Kevin Wang, Yuedong Wang, Peter Kotanko
Frontiers in Nephrology, 2022.
This study evaluates immune responses among hemodialysis patients following COVID-19 infection and three doses of the mRNA-1273 vaccine, providing evidence on neutralizing antibody responses in a high-risk population.
Reproducing Kernel Hilbert Space
Reproducing Kernel Hilbert Space
Jiahui Yu, Kevin Wang, Anna Liu, Yuedong Wang
Wiley StatsRef: Statistics Reference Online, 2022.
A reference article introducing the theory of reproducing kernel Hilbert spaces (RKHS), including their mathematical foundations and applications in modern statistical learning and machine learning.
Talks and Awards
- Deep Generative Models: Complexity, Dimensionality and Approximation
ICSA 2025 Student Paper Award Session
2025
Presented theoretical results on approximation, latent dimensionality, and complexity trade-offs in deep generative models on manifolds.
Biostatistics 75th Anniversary Conference & Celebration
University of North Carolina at Chapel Hill
2024
Invited presentation on theoretical foundations of deep generative models, including approximation guarantees for distributions supported on low-dimensional manifolds.
Page updated
Google Sites
Report abuse