Constraining Deep Generative Models with Neurosymbolic Approach

Zhe Zeng
UVA CS

Time: 2025-10-29, 12:00 - 13:00 ET
Location: Rice 540 and Zoom

Abstract Deep Generative Models (DGMs) emerged as powerful tools for generating complex, high-dimensional data and capturing the distributions that characterize them. Despite their impressive capabilities, they struggle when the data is subject to constraints that encode background knowledge; they are not able to learn even the simple background knowledge and thus can fail to generate sample compliant with constraints. Existing solutions to these problems have primarily relied on heuristic methods, often disregarding the underlying data distribution. As a result, incorporating constraints into DGMs remains a critical challenge. In this talk, I will first demonstrate the current limitations using the fixed-sum constraint as examples. I will then present recent advances in integrating constraints into DGMs to capture relevant problem structure and generate compliant and realistic samples. Finally, I will discuss some potential future research directions.

Bio: Zhe Zeng is an Assistant Professor in the Department of Computer Science at University of Virginia. Her research centers on neurosymbolic AI and probabilistic machine learning to build models that are inherently compliant with domain-specific constraints and capable of leveraging the background knowledge these constraints provide, aiming to achieve reliable, interpretable and trustworthy AI and aid scientific discoveries. Previously, Zhe was a faculty fellow in the Computer Science Department at New York University. She obtained her Ph.D. in Computer Science at the University of California, Los Angeles. She received the Amazon Doctoral Student Fellowship in 2022 and the NEC Student Research Fellowship in 2021, and was selected for UVA Engineering Rising Scholars in 2025, and the Rising Stars in EECS in 2023.