Minge Xie
Rutgers University-New Brunswick, NJ·Mathematics & Statistics
When a machine learning model makes a prediction, how much should we trust it? Many modern tools, such as deep neural networks and random forests, can be accurate yet act as "black boxes" that are hard to interpret.
- Fields
Mathematics & Statistics
Computer Science
- Topics
statistical inference
uncertainty quantification
machine learning
- Grants
- 2 known grants, through Jul 2028
What they work on
- Develop PCSI (Predictability-Computability-Stability Inference) for binary and multi-class classification, screening many algorithms by predictive performance and using bootstrap resampling to improve uncertainty estimates, robustness and accuracy.
- Design computationally feasible PCSI methods for deep learning and large language models, using data perturbations and calibration to make predictions more stable, interpretable and accurate.
- Establish theory showing that PCSI and PCS-guided ensemble learning remain valid even when the models are wrong, and can beat conformal inference under reasonable conditions.
- Quantify uncertainty in decision trees and random forests, models that combine many simple yes/no splitting rules, so that their predictions come with performance-guaranteed inference.
- Develop performance-guaranteed methods to discover deep neural network architectures and to perform statistical inference for generative graphical networks.
- Build general theory for irregular inference problems involving discrete or non-numerical parameters and non-numerical data, where standard large-sample statistical theory does not apply.
More about the research
Standard statistics often assumes the data come from one specific probability model, and that assumption often fails for complex modern data. The result can be misleading conclusions and studies that others cannot reproduce. These problems matter most in areas like health care. Across two projects, the lab is building statistical theory and methods to measure uncertainty in machine learning and to make its results interpretable, reliable and reproducible.
Why it matters
Data science and AI now guide decisions in science, health care, technology and national security. Better ways to measure and report uncertainty could help scientists, clinicians and policymakers judge when model-based conclusions are trustworthy. This could reduce misleading or non-reproducible findings. More interpretable machine learning could also strengthen trust in AI used in medical settings.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Collaborative Research: Performance Guaranteed Statistical Learning with Multiple Classes of Models (guided by PCS)
NSF · Aug 2025 – Jul 2028
Unravel machine learning blackboxes -- A general, effective and performance-guaranteed statistical framework for complex and irregular inference problems in data science
NSF · Jul 2023 – Jun 2027
Closest labs
Matias Cattaneo
Princeton University, NJ·Mathematics & Statistics
Demonstrate the empirical and theoretical limitations of binned regression models applied to mixed-frequency panel data in climate economics, showing when they can lead to incorrect conclusions.
Daniel Nordman
Iowa State University, IA·Mathematics & Statistics
Develop the blended bootstrap, a general strategy that merges the strengths of subsampling and the bootstrap, two different ways of re-using data to measure uncertainty, for more accurate inference on complex data.
Yao Zheng
University of Connecticut, CT·Mathematics & Statistics·Early-career
Develop recurrent neural network (RNN) models for high-dimensional time series that stay interpretable through Granger causality, a test of whether one series predicts another, to balance accuracy and interpretability.
The school
Rutgers University-New Brunswick: 57 labs on PI Pages, $837M a year on research, 713 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.