Edgar Dobriban

Assistant Professor of Statistics and Data Science, with secondary appointment in Computer and Information Science

Contact Information

Primary Email:
dobriban@wharton.upenn.edu

office Address:
305 Academic Research Building
265 South 37th Street
Philadelphia, PA 19104

Research Interests: Statistics and machine learning

Overview

A postdoctoral research fellow position is available to work on adaptive and adversarial machine learning, with connections to developmental psychology. Please see ad and send e-mail to Edgar.

New class in Spring 2022: Topics in Modern Statistical Learning (STAT-991), surveying uncertainty quantification for machine learning. See the Github page.

Research interests:

the efficient statistical analysis of “big data” using advanced tools, such as those from random matrix theory
- dimension reduction, PCA: [1], [2], [3], [4], [5], [6]
- multiple testing: [1], [2], [3]
- high-dimensional regression: [1], [2]
- invariance-based randomization tests
- statistically efficient prediction sets
the foundations of modern machine learning, including deep learning
- uncertainty quantification: [1], [2], [3], [4]
- sketching and random projections, [1], [2], [3], [4], [5]
- data augmentation and symmetry: [1], [2], [3], [4]
- distributed learning: [1], [2], [3]
- (stochastic) gradient descent and flow: [1], [2]
- fairness: [1], [2]
- overparametrization
- retraining of ML models
- reinforcement learning inspired by child-like learning
- adversarial robustness: [1], [2]
- weight normalization
in addition, we occasionally work on important applications and methods, such as
- genomics
- group testing for COVID-19

The group is always looking to expand. We are recruiting PhD students at Penn to work on problems in statistics and machine learning. PhD applicants interested to work with me should mention this on their application. Please apply through the departments of Statistics & Data Science, Computer and Information Science, and the AMCS program, as it gives higher chances for admission.

Education (cv):

PhD in Statistics, Stanford University, 2017. Advisor: David Donoho
BA in Mathematics, Princeton University, 2012.

Research

Talk slides: GitHub. Google Scholar.

Sangdon Park, Edgar Dobriban, Osbert Bastani, Insup Lee PAC Prediction Sets for Meta-Learning.
Druv Pai, Michael Psenka, Chih-Yuan Chiu, Manxi Wu, Edgar Dobriban, Yi Ma Pursuit of a Discriminative Representation for Multiple Subspaces via Sequential Games.
Yinshuang Xu, Jiahui Lei, Edgar Dobriban, Kostas Daniilidis (2022), Unified Fourier-based Kernel and Nonlinearity Design for Equivariant Networks on Homogeneous Spaces,.
Souradeep Dutta, Yahan Yang, Elena Bernardis, Edgar Dobriban, Insup Lee, Memory Classifiers: Two-stage Classification for Robustness in Machine Learning.
Xinmeng Huang, Donghwan Lee, Edgar Dobriban, Seyed Hamed Hassani, Collaborative Learning of Distributions under Heterogeneity and Communication Constraints.
Shuo Li, Xiayan Ji, Edgar Dobriban, Oleg Sokolsky, Insup Lee PAC-Wrap: Semi-Supervised PAC Anomaly Detection.
Xianli Zeng, Edgar Dobriban, Guang Cheng Fair Bayes-Optimal Classifiers Under Predictive Parity.
Evangelos Chatzipantazis, Stefanos Pertigkiozoglou, Edgar Dobriban, Kostas Daniilidis, SE(3)-Equivariant Attention Networks for Shape Reconstruction in Function Space.
Hongxiang Qiu, Edgar Dobriban, Eric Tchetgen Tchetgen (Draft), Distribution-free Prediction Sets Adaptive to Unknown Covariate Shift.
Abstract: Predicting sets of outcomes -- instead of unique outcomes -- is a promising solution to uncertainty quantification in statistical learning. Despite a rich literature on constructing prediction sets with statistical guarantees, adapting to unknown covariate shift -- a prevalent issue in practice -- poses a serious challenge and has yet to be solved. In the framework of semiparametric statistics, we can view the covariate shift as a nuisance parameter. In this paper, we propose a novel flexible distribution-free method, PredSet-1Step, to construct prediction sets that can efficiently adapt to unknown covariate shift. PredSet-1Step relies on a one-step correction of the plug-in estimator of coverage error. We theoretically show that our methods are asymptotically probably approximately correct (PAC), having low coverage error with high confidence for large samples. PredSet-1Step may also be used to construct asymptotically risk-controlling prediction sets. We illustrate that our method has good coverage in a number of experiments and by analyzing a data set concerning HIV risk prediction in a South African cohort study. In experiments without covariate shift, PredSet-1Step performs similarly to inductive conformal prediction, which has finite-sample PAC properties. Thus, PredSet-1Step may be used in the common scenario if the user suspects -- but may not be certain -- that covariate shift is present, and does not know the form of the shift. Our theory hinges on a new bound for the convergence rate of Wald confidence interval coverage for general asymptotically linear estimators. This is a technical tool of independent interest.
Donghwan Lee, Xinmeng Huang, Seyed Hamed Hassani, Edgar Dobriban, T-Cal: An optimal test for the calibration of predictive models.

Awards and Honors

Junior research award, ICSA 2022 China Conference, 2022
NSF CAREER Award, 2021
NSF-Simons Award on Mathematical and Scientific Foundations of Deep Learning, co-PI, 2020
NSF Harnessing the Data Revolution TRIPODS Award, co-PI, 2019
T.W. Anderson Theory of Statistics Dissertation Award, Department of Statistics, Stanford University, 2017
Howard Hughes Medical Institute International Student Graduate Research Fellowship, 2015
Stanford Department of Statistics Teaching Award, 2013
Middleton Miller ’29 Prize for best independent work in mathematics, Princeton University, 2012

Miscellaneous

This page has links to methods from my papers. Feel free to contact me if you are interested to use them.

ePCA: github

The ePCA method for principal component analysis of exponential family data, e.g. Poisson-modeled count data. (with L.T. Liu);

EigenEdge: github

Methods for working with large random data matrices, including

Computing eigenvalue distributions of covariance matrices (general Marchenko-Pastur distributions).
Optimal statistics for testing in principal component analysis.
Tools for spiked covariance models: spike and cosine descriptors, optimal shrinkers.

pweight : github R. github Matlab

P-value weighting techniques for multiple hypothesis testing. These can improve power in multiple testing, if there is prior information about the individual effect sizes. Includes the iGWAS method for Genome-Wide Association Studies.

Activity

Latest Research

Sangdon Park, Edgar Dobriban, Osbert Bastani, Insup Lee PAC Prediction Sets for Meta-Learning.

All Research