Nominate a speaker »
**Seminars will be held In Person unless otherwise noted**
Fall 2025
Seminar One
September 11, 2025
Griffin Floyd Hall, Room 100
Martin Slawski
University of Virginia
Title:
Linked Data Analysis without Linkage
Abstract:
Record Linkage (RL) is widely regarded as the gold standard for leveraging information scattered across multiple files. At the same time, RL can be tedious, error-prone, and privacy-invasive, which has hampered a more widespread and streamlined adoption. In this talk, I introduce Unlinked Data Analysis as a novel paradigm enabling inference for associations of variables across files, without linking them.
In the first part, I discuss the task of learning a map between d-dimensional inputs and d-dimensional noisy outputs, without observing (input, output)-pairs, but only separate unordered lists of inputs and outputs. I show that this problem can be solved for Brenier maps and present a practical algorithm for denoising based on the Kiefer-Wolfowitz nonparametric maximum likelihood estimator (NPMLE) and techniques from optimal transport. I provide upper bounds on the resulting mean squared denoising error under (i) Gaussian noise and (ii) noise from a certain class of elliptic distributions.
In the second part, I give an overview of linear regression in the unlinked setting, with a particular focus on existing results and open problems related to identifiability.
Seminar Two
September 25, 2025
Zoom ID: https://ufl.zoom.us/j/93709716185
Jason Altschuler
University of Pennsylvania
Title:
Shifted Divergences for sampling, privacy, and beyond
Abstract:
Shifted divergences provide a principled way of making information-theoretic divergences (e.g., KL) geometrically aware via optimal transport smoothing. In this talk, I will argue that shifted divergences provide a powerful approach towards unifying central problems in optimization, sampling, privacy, functional inequalities, and beyond. For concreteness, I will describe these connections by mentioning several recent highlights, focusing on the first:
(1) Characterizing the mixing time of the Langevin Algorithm to its stationary distribution for log-concave sampling.
(2) The fastest high-accuracy algorithm for sampling from log-concave distributions.
(3) A positive answer to the acceleration conjecture in sampling.
(4) Characterizing the differential privacy of Noisy-SGD, the standard algorithm for private convex optimization.
(5) Tight shift-Harnack inequalities and simple proofs of Wang’s celebrated dimension-free Harnack inequalities. A recurring theme is a certain notion of algorithmic stability, and the central technique for establishing this is shifted divergences (or its more powerful abstraction, the Shifted Composition Rule).
Based on joint works with Kunal Talwar, with Sinho Chewi, and with Jinho Bok and Weijie Su.
Seminar Three
October 9, 2025
Griffin-Floyd Hall, Room 100
Krishna Balasubramanian
University of California-Davis
Title:
Riemannian Proximal Sampler for High-accuracy Sampling on Manifolds
Abstract:
Sampling from densities defined on Riemannian manifolds is central to Bayesian inference, generative modeling, and differential privacy. We introduce the Riemannian Proximal Sampler (RPS), whose efficiency hinges on two oracles: Manifold Brownian Increments and the Riemannian Heat Kernel. We establish high-accuracy sampling guarantees for the Riemannian Proximal Sampler, showing that generating samples with ε-accuracy requires O(log(1/ε)) iterations in Kullback-Leibler divergence assuming access to exact oracles and O(log^2(1/ε)) iterations in the total variation metric assuming access to sufficiently accurate inexact oracles. Furthermore, we present two practical implementations of these oracles by leveraging heat-kernel truncation and Varadhan’s asymptotics, respectively. In the latter case, we interpret the Riemannian Proximal Sampler as a discretization of the entropy-regularized Riemannian Proximal Point Method on the associated Wasserstein space. We will discuss numerical results that illustrate the effectiveness of the proposed methodology.
Seminar Four
October 23, 2025
Griffin-Floyd Hall, Room 100
Sanjay Chaudhuri
University of Nebraska
Title:
On an Empirical Likelihood-Based Solution to Approximate Bayesian Computation Problem
Abstract:
For many complex models studied in natural, engineering, and environmental sciences, it is nearly impossible to specify a likelihood for the observed data. Approximate Bayesian Computation (ABC) methods try to estimate such model parameters only by comparing the given observation and some replicates generated from the model for various input parameter values. No explicit relationship between the parameters and the data is postulated. In this article, we propose an empirical likelihood (EL) based solution to the ABC problem. By construction, our method is based on an interpretable likelihood (i.e. the EL) which is computed using estimating equations completely specified by the observed and the replicated data and a few well-chosen summary statistics. The proposed method can be justified through information projections on a specified class of densities. We further show that the posterior is consistent and discuss several of its favourable large sample and large replication properties. Illustrative examples from various real-life applications will also be presented.
This work is joint with Subhroshekhar Ghosh and Pham Thi Kim Cuc, all from the National University of Singapore.
Seminar Five-Challis Lecture-Day One
October 28, 2025
Reitz Union, Room G320
Nancy Reid
University of Toronto
Title:
TBD
Abstract:
TBD
Seminar Six-Challis Lecture-Day Two
October 29, 2025
Reitz Union, Room G320
Nancy Reid
University of Toronto
Title:
TBD
Abstract:
TBD
Seminar Seven
November 6, 2025
Griffin-Floyd Hall, Room 100
Didong Li
UNC Chapel Hill
Title:
TBD
Abstract:
TBD
Seminar Eight
November 13, 2025
Griffin-Floyd Hall, Room 100
Nandita Mitra
University of Pennsylvania
Title:
TBD
Abstract:
TBD