Upcoming Events
PhD Defense | Representation and Statistical Properties of Deep Neural Networks and Diffusion Models on Structured Data
Title: Representation and Statistical Properties of Deep Neural Networks and Diffusion Models on Structured Data
Date: Monday, November 23, 2026
Time: 3:30 PM – 5:30 PM ET
Location: Online via Microsoft Teams
Zixuan Zhang
Machine Learning PhD Student
H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology
Committee
1. Dr. Tuo Zhao (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
2. Dr. Shihao Yang, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
3. Dr. Sen Na, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
4. Dr. Wenjing Liao, School of Mathematics, Georgia Institute of Technology
5. Dr. Minshuo Chen, Department of Industrial Engineering and Management Sciences, Northwestern University
Abstract
Deep learning succeeds on high-dimensional data far beyond what classical statistical theory predicts: neural networks learn images from a modest number of samples, and diffusion models learn their distributions. A common explanation is that real data are structured and that these models adapt to the structure. This dissertation makes the explanation precise by identifying an effective complexity of the data distribution and proving that both the network size needed to represent the target and the sample complexity of learning it are governed by this complexity.
The main part of the dissertation applies this principle to distribution estimation with diffusion models. For score-based diffusion models on data supported on a manifold, we establish approximation and estimation theories in which the network size and the sample complexity are governed by the intrinsic dimension, and we obtain minimax optimal rates for distribution estimation in Wasserstein distance. The analysis also quantifies the effect of data geometry: higher curvature of the manifold leads to larger approximation and estimation errors. For discrete diffusion models on sequence data, we develop a framework for generalization under arbitrary forward corruption kernels and prove that the sample complexity of masked corruption scales with the effective support of the data whereas that of uniform corruption scales with the full state space, the first result that separates the statistical efficiency of different forward processes. Together, these results show that deep neural networks and diffusion models adapt to the effective complexity of structured data and advance the statistical foundations of deep learning and generative modeling.
Event Details
-
Online
Media Contact
EVENTS BY SCHOOL & CENTER
School of Computational Science and Engineering
School of Interactive Computing
School of Cybersecurity and Privacy
Algorithms and Randomness Center (ARC)
Center for 21st Century Universities (C21U)
Center for Deliberate Innovation (CDI)
Center for Experimental Research in Computer Systems (CERCS)
Center for Research into Novel Computing Hierarchies (CRNCH)
Constellations Center for Equity in Computing
Institute for People and Technology (IPAT)
Institute for Robotics and Intelligent Machines (IRIM)