Kaizheng Wang
Columbia University, NY·Mathematics & Statistics·Early-career
Modern data rarely come in one clean, trustworthy batch. Data may be synthetic output from a generative AI model whose errors are hard to detect.
- Fields
Mathematics & Statistics
Computer Science
- Topics
generative AI
uncertainty quantification
synthetic data
- Grants
- 2 known grants, through Aug 2031
What they work on
- Develop methods to measure the overall fidelity of black-box generative AI models using effective sample size, an interpretable estimate of how much reliable information synthetic data contain.
- Create local, context-aware uncertainty estimates showing when a generative model is reliable for specific inputs, subpopulations, and shifting environments, and when its outputs need caution.
- Design statistically guided training strategies for generative models used in counterfactual analysis and choice modeling, including AI-assisted articulation of human needs and knowledge distillation.
- Build transfer learning methods that integrate samples from multiple heterogeneous sources and adapt to temporal distribution shifts in evolving data streams, with theoretical guarantees under minimal assumptions.
- Develop approaches for learning latent structures by combining multiple data views of the same subjects, such as social network information with individual user profiles.
More about the research
Or they may be collected from many sources, at different times, and in different formats. Both projects ask the same statistical question: how can we draw reliable conclusions from imperfect data, and how can we tell how much each piece of data can be trusted? The work develops methods, theory, and open-source software for these problems.
Why it matters
Organizations increasingly rely on synthetic data and on data pooled from many sources, but often cannot tell when results are trustworthy. These methods could help researchers, educators, and businesses judge when AI-generated data can be relied on. Examples include safely testing educational tools on simulated student populations, and judging autonomous vehicle safety in a new city when local test data are scarce.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Yinqiu He
University of Wisconsin-Madison, WI·Mathematics & Statistics
Develop methods with theoretical guarantees for estimating low-dimensional network embeddings and quantifying their uncertainty under general network models with relaxed assumptions.
Xinyi Li
Clemson University, SC·Mathematics & Statistics
Develop interpretable statistical learning methods for single-stage individualized treatment rules that combine medical imaging, genomics and clinical variables, with valid uncertainty quantification and statistical inference.
Cong Ma
University of Chicago, IL·Mathematics & Statistics·Early-career
Develop robust ranking algorithms for pairwise comparison data that tolerate ill-conditioned sampling of which pairs are compared and relax parametric modeling assumptions.
The school
Columbia University: 89 labs on PI Pages, $1.4B a year on research, 625 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.