Time: 4:30-5:30 p.m.
Date: Wednesday, November 12, 2025
Speaker: Ziv Goldfeld,Associate Professor, School of Electrical & Computer Engineering, Cornell University
Title: Robust & Geometry-Aware Distribution Estimation via Optimal Transport
Abstract: Modern machine learning relies on large, high-dimensional datasets with rich geometric structure. As their reach expands, so does the risk posed by data poisoning attacks and incorrect modeling assumptions. We develop robust and geometry-aware learning algorithms subject to statistical and computational guarantees via optimal transport (OT)—a versatile framework amenable to analysis and algorithms. We consider decision-making in an adversarial environment, where an ε-fraction of samples from an unknown distribution are arbitrarily modified (global corruptions) and the remaining perturbations have average magnitude bounded by ρ (local corruptions). From such corrupted data, we study the statistical and computational landscape of distribution estimation under the Wasserstein metric. We propose an estimator based on a Wasserstein projection under a partial transportation constraint and show that it is the first estimator to achieve minimax optimality simultaneously across all corruption budgets ε,ρ>0. Computationally, we approximate the projection via a multivariate extension of iterative filtering for robust mean estimation, yielding an efficient algorithm. We also present a heuristic neural implementation via adversarial training, rooted in a new duality theory for partial OT.
This talk is based on joint work with Sloan Nietert, Soroosh Shafiee, and Rachel Cummings.
Bio: Ziv Goldfeld is an Associate Professor at the School of Electrical and Computer Engineering, and a graduate field member of Computer Science, Statistics and Data Science, Operations Research and Information Engineering, and the Center of Applied Mathematics at Cornell University.