Statistics Seminar - Fall 2026
Seminars are held on Thursdays from 1:00 - 2:00pm on Zoom unless otherwise noted. For access information, please contact the Math Department.
For questions about the seminar schedule, please contact Yan Sun
Link for Zoom: https://njit-edu.zoom.us/j/92467097971?pwd=3jattprgEBHkXWwibfnu5pKSb388rg.1
Meeting ID and passcode will be provided in announcement.
*******************************************************************************************************
October 1
Buxin Su, University of Pennsylvania
Statistics for AI and Society: Evaluating AI Research and Improving Public Data
In the first part, I show how authors’ knowledge of their own work, specifically their rankings of their own papers, can help improve peer review at AI conferences. We study the isotonic mechanism, which calibrates review scores according to authors’ rankings, and present results from experiments conducted from 2023 to 2025 at ICML, one of the largest AI conferences. We find that authors’ self-rankings predict their papers’ future citation impact even after accounting for review scores. When combined with review scores, these rankings also improve estimates of paper quality. In 2026, author self-rankings became part of ICML’s review policy, and I discuss what we learned from its implementation. In the second part, I present practical methods for improving the utility of public social data. Our methods can reduce noise variance in U.S. Decennial Census data by about 20% while meeting a specified privacy requirement. This improvement is particularly beneficial for minority populations and communities.
This talk is based on joint work with Weijie Su, Bingxin Zhao, and Jianqing Fan. The research has been featured in several publications, including [Nature Index](https://www.nature.com/articles/d41586-026-00059-0), [Nature Computational Science News & Views](https://www.nature.com/articles/s43588-026-01046-1), and [New Scientist](https://www.newscientist.com/article/2454095-simple-fix-could-make-us-census-more-accurate-but-just-as-private/).
Homepage: https://buxinsu.github.io/
*******************************************************************************************************
October 22
Frank Shih, Memorial Sloan Kettering Cancer Center
Deep Neural Networks as Statistical Models: Uncertainty Quantification Through Extended Fiducial Inference
Deep neural networks have transformed predictive modeling, yet extending statistical inference to these flexible models remains challenging. Even when a likelihood is readily available, parameter nonidentifiability and the difficulty of characterizing the sampling distributions of estimators complicate uncertainty quantification. Bayesian approaches face additional challenges in prior specification and posterior computation. Scientific and engineering applications therefore motivate a framework that connects flexible modeling with inference for underlying functional relationships, scientific parameters, and future observations.
This talk introduces extended fiducial inference (EFI), which approaches statistical inference from an equation-solving point of view. Through an energy-based formulation of data-generating equations, EFI explores the set of parameter and latent noise values that satisfy the equations for the observed data. It then propagates uncertainty in the latent noise to the inferential targets, without requiring a prior on the target parameters. Applications to deep neural networks and scientific modeling illustrate its potential to integrate flexible prediction, physical constraints, and uncertainty quantification within a common framework.
Homepage: https://scholar.google.com/citations?user=yBogbycAAAAJ&hl=en
*******************************************************************************************************
November 5
Sanat K. Sarkar, Temple University
Title Forthcoming
Abstract Forthcoming
Homepage: https://www.fox.temple.edu/directory/sanat-k-sarkar-sanat
*******************************************************************************************************
November 12
Tareq Aldirawi, NJIT
Title Forthcoming
Abstract Forthcoming
Homepage:
*******************************************************************************************************
November 19
Dongyuan Song, University of Connecticut
Title Forthcoming
Abstract Forthcoming
Homepage: https://facultydirectory.uchc.edu/profile?profileId=Song-Dongyuan
*******************************************************************************************************
December 3
Jing Wang, University of Pennsylvania
Title Forthcoming
Abstract Forthcoming
Homepage: https://maayawang.github.io/mysite/
*******************************************************************************************************
December 10
Tianlu Xue, NJIT
Title Forthcoming
Abstract Forthcoming
Homepage:
*******************************************************************************************************
Last updated: October 19, 2026