Logistics
- When: Thursdays 12-1 pm
- Where: Amy Gutmann Hall 414
- Organizers: Pratik Chaudhari, Meena Jagadeesan, Enric Boix-Adsera, Rene Vidal
- Administrative coordinator: Sonia Castro (soniacr@upenn.edu)
- Mailing list: subscribe here to “folds-seminar”
About
What This seminar features leading experts in optimization, learning, and data science. Topics span algorithms, complexity, modeling, applications, and mathematical underpinnings.
Why Foundational advances in these fields are increasingly intertwined. This seminar serves as a university-wide hub to bring together the many communities across UPenn interested in optimization, learning, and data science: the Department of Statistics and Data Science, Electrical Engineering, Computer Science, Applied Mathematics, Economics, Wharton OID, and others. To help promote this internal interaction, several speakers will be from UPenn.
Funding We are grateful to the IDEAS Center for Innovation in Data Engineering and Science, Penn AI, and the Wharton Department of Statistics and Data Science.
Previous seminars
- Spring 2026
- Fall 2025
- Previously evolved from the UPenn Optimization seminar: Fall 2023, Spring 2024, Fall 2024, Spring 2025
Fall 2026
| Date | Speaker |
|---|---|
| Sep 3 | Shanyin Tong |
| Sep 10 | Morgane Austern |
| Sep 17 | Mateo Díaz |
| Sep 24 | Pablo Gonzalez-Camara |
| Oct 1 | Fall Break |
| Oct 8 | Nicolas Loizou |
| Oct 15 | Lachlan MacDonald |
| Oct 29 | Meena Jagadeesan (in AGH 306) |
| Nov 5 | Jonathan Niles-Weed |
| Nov 12 | Cynthia Rush |
| Nov 19 | Daniel Robinson |
Spring 2027
| Date | Speaker |
|---|---|
| March 18 | Carlos Fernandez-Granda |
| March 25 | Eric Bradlow |
| April 1 | Claire Donnat |
| April 8 | Ying Jin |
| April 15 | Enric Boix |
| April 22 | Shipra Agarwal |
Talk Abstracts
Date: Sep 3, 2026
Title: Large Deviations for Rare-Event Estimation and Control
Abstract: Rare and extreme events, such as natural disasters, cascading failures, and accidents in autonomous systems, can have severe consequences but are inherently data-scarce: precisely because they occur infrequently, there are often too few observations to reliably learn their statistics directly from data. This creates a fundamental challenge for data-driven prediction and decision-making, particularly when the underlying systems involve high-dimensional uncertainty and expensive physical models. In this talk, I will present a computational framework based on large deviation theory (LDT) for rare-event estimation and control. LDT connects the probability of a rare event to a deterministic optimization problem over the uncertain parameters, whose solution identifies the most likely mechanism leading to the event. Building on this connection, I will discuss sampling-free approximations for rare-event probabilities, LDT-informed importance sampling algorithms, and optimization under rare chance constraints. These approaches exploit the geometry of the rare-event set and information from the associated optimization problem to substantially reduce the computational cost of both probability estimation and risk-aware control.
I will illustrate these methods through applications to physical and engineered systems, with a particular focus on traffic systems involving autonomous vehicles. These developments form part of RareDT, a broader effort to develop digital twins that go beyond predicting typical system behavior and instead incorporate the quantification and control of rare but consequential events.
Date: Sep 10, 2026
Title: Dependence and degeneracy create: Multiple descent in overparameterized models
Abstract: Over-parameterized linear regression has been widely studied over the last decade. However, most existing works assume that the covariates are independent and that their covariance matrices are non-degenerate. In this talk, we will relax both assumptions and derive deterministic equivalents for the prediction risk in a vanishing-ridge regime. We show that degeneracy of the covariance matrices and dependence can lead to multiple descent, and characterize where the corresponding peaks can occur. Our proofs use a novel graph representation of the variance profile. We show that maximum matchings and the Dulmage–Mendelsohn decomposition of the associated bipartite graph identify the configurations at which the variance becomes singular.
Date: Sep 17, 2026
Title: Leveraging Structure for Faster Algorithms in Optimization and Diffusion
Abstract: Large-scale iterative methods drive modern AI, yet their theoretical foundations often lag behind their empirical success. We argue that bridging this gap requires identifying the inherent problem structure that enables these algorithms to perform well. This talk instantiates this principle across two domains: optimization and generative modeling.
First, we derive new theoretical guarantees for the Levenberg–Morrison-Marquardt method. Although this method is ubiquitous in settings that demand highly accurate solutions—for instance, when training physics-informed neural networks for scientific discovery—classical guarantees do not explain its strong empirical performance in modern overparameterized, ill-conditioned regimes. By reframing it through the lens of composite optimization, we uncover geometric conditions that ensure fast convergence even in these challenging modern regimes.
Second, we introduce Proximal Diffusion Models (PDM). While standard diffusion models rely on score-matching and forward discretization, we demonstrate that a backward discretization using proximal maps offers significant theoretical and practical advantages. Under mild conditions, we prove that PDM achieves ε-accuracy in KL-divergence within $\widetilde{O}(d/\sqrt{\varepsilon})$ steps and empirically demonstrate that it outperforms conventional methods using fewer sampling iterations.
Bio: Mateo Díaz is an Assistant Professor in the Department of Statistics and the Data Science Institute at the University of Chicago. His research lies at the intersection of continuous optimization, geometry, and statistics. He develops mathematical foundations and scalable algorithms for problems arising in data science, machine learning, and signal processing. Before joining the University of Chicago, he was an Assistant Professor of Applied Mathematics and Statistics at Johns Hopkins University. He previously spent two years as a postdoctoral scholar at Caltech. Mateo received his Ph.D. in Applied Mathematics from Cornell University in 2021 and earned bachelor’s degrees in Mathematics and in Systems and Computing Engineering, as well as a master’s degree in Mathematics, from Universidad de los Andes in Colombia. His work has been recognized with the Beale–Orchard-Hays Prize, an NSF CAREER Award, and a Sloan Research Fellowship in Mathematics.
Date: Sep 24, 2026
Title: Metric Geometry and the Study of Biological Shape Across Scales
Abstract: The systematic study of biological shape has been central to major advances in biology, from the development of the neuron doctrine in the 19th century to the discovery of the molecular basis of sickle cell disease and the structures of DNA and hemoglobin. Across scales, biological shape encodes fundamental information about biological function and organization. Over the past two decades, advances in microscopy and structural biology have enabled increasingly rich and precise measurements of biological structures at the cellular, subcellular, and molecular scales. These developments create a need for rigorous quantitative methods for representing, comparing, and analyzing biological shapes, especially when the objects under study are highly heterogeneous. In this talk, I will discuss recent applications of ideas from metric geometry, and particularly Gromov-Wasserstein couplings, to these problems. We will focus on three settings: comparing cellular morphologies across heterogeneous and morphologically complex populations, such as neurons and glia; analyzing subcellular organization in ways that are robust to variation in cellular morphology; and comparing three-dimensional protein structures. We will also discuss how these methods can be naturally combined with metric-learning neural networks to achieve the scalability of deep learning while retaining interpretability and generalizability. Together, these results illustrate how metric geometry can provide a rigorous and flexible quantitative framework for studying biological shape across scales.
Bio Pablo G. Cámara is an Associate Professor of Genetics at the University of Pennsylvania and a faculty member of the Institute for Biomedical Informatics and the AI2D Center for AI and Data Science for Integrated Diagnostics. His research focuses on the cellular and molecular organization of the brain and brain tumors, utilizing mathematical principles to analyze high-dimensional omics and imaging data. He received a Ph.D. in Theoretical Physics from Universidad Autónoma de Madrid and continued his postdoctoral work at École Polytechnique (France), the European Organization for Nuclear Research (CERN, Switzerland), and the University of Barcelona. Fascinated by open fundamental questions in biomedicine, he shifted his focus to quantitative biology in 2014 as a postdoctoral fellow at Columbia University and the Institute for Advanced Study in Princeton. He joined the University of Pennsylvania as a faculty member in 2018, where he is developing geometry- and topology-based algorithms for integrating and analyzing single-cell omics, cytometry, and imaging data, and using them to elucidate the cellular ecosystem and oncogenic pathways of glioma, and to characterize the functional properties of CAR-T cell immunotherapies.
Date: Oct 8, 2026
Title: Extragradient Methods for Modern Machine Learning Settings: New Convergence Guarantees, Step-Size Rules, and Stochastic Variants
Abstract: Extragradient methods are a fundamental class of algorithms for solving min-max optimization problems and variational inequalities. While the classical theory is largely developed under smoothness and other relatively restrictive assumptions, many problems arising in modern machine learning call for analysis in weaker regularity regimes and in stochastic large-scale settings. In this talk, we present new convergence results for deterministic and stochastic extragradient methods beyond the classical framework. In particular, we establish convergence guarantees under the (L0, L1)-Lipschitz condition and derive new step-size rules that expand the range of provably convergent regimes. We also introduce Polyak-type step sizes for deterministic and stochastic extragradient methods, leading to adaptive variants with favorable theoretical properties and practical performance. Our results focus primarily on monotone problems, with extensions to selected structured non-monotone settings. We conclude with numerical experiments that illustrate the theory and the empirical behavior of the proposed methods.
Date: Oct 15, 2026
Title: Dynamics of large-step gradient descent near flat minima
Abstract: Classically, convergence guarantees for gradient descent (GD) require that the step size be sufficiently small compared to the local curvature (“sharpness”) of the objective. This “small-step” condition is routinely violated in the practical training of deep learning models, wherein the favoured “large-step” GD is observed to converge with intriguing oscillatory dynamics and an implicit bias toward solutions of low curvature known as “flat minima”. In this talk, I will present a new geometric perspective on these phenomena. In particular, I will indicate how the loss geometry of certain multilayer models enables the proof of a new normal form for large-step GD near the global minima manifold. In this normal form, large-step GD is revealed to act as Riemannian GD on the objective sharpness along the solution manifold coupled to a flip bifurcation in the directions orthogonal to the solution manifold, enabling new local convergence theorems for large-step GD near flat minima.
Bio: Lachlan MacDonald is a postdoc in the Vidal Lab at the University of Pennsylvania studying the mathematical foundations of deep learning. He received his PhD in pure mathematics from the University of Wollongong in 2019 and held postdoc positions in pure mathematics at the Australian National University and the University of Adelaide. He shifted his focus to the theory of deep learning in 2021, holding postdoc positions at the University of Adelaide and Johns Hopkins University before joining UPenn in 2024. His research leverages differential geometry and dynamical systems theory to understand the nonlinear dynamics and implicit bias of neural network training.
Date: Oct 29, 2026
Title: Strategic Behavior in Machine Learning: Case Studies in LLM Detection and Medicare Advantage
Abstract: When ML models make high-stakes decisions, these models create incentives for agents to strategically change their behavior. In this talk, we discuss how addressing these incentives requires application-centered models of strategic behavior. We focus on two case studies. The first is LLM detection, where users jointly decide how much to use an LLM and how much to post-process features used for detection. The second is Medicare Advantage payments, where decision-makers rely on coarse levers such as feature selection. Our results uncover new failure modes and identify feature-based interventions that account for these application-specific complexities.
Bio: Meena Jagadeesan is an Assistant Professor in the Department of Computer and Information Science at the University of Pennsylvania. Her research focuses on machine learning ecosystems, and she specifically works on safety, incentive, and design problems that emerge from multi-agent interactions. Before joining Penn, she received a PhD from UC Berkeley and did a postdoc at Stanford.
Date: Nov 5, 2026
Title: Wasserstein Least Squares: From Snapshots to Trajectories
Abstract: This talk will present Wasserstein least squares, a regression method for vector-valued covariates and distribution-valued responses with a direct interpretation in terms of random variables. We will show that it is the canonical extension of Euclidean least squares to the space of probability distributions, and that, surprisingly, estimation is possible at the n^(−1/2) rate. We will then explore extensions to kernel ridge regression and to trajectory inference. These extensions come with statistical guarantees and polynomial-time algorithms, and yield more regular and interpretable trajectories than existing approaches. Applications include large-scale demographic data from the RAND Health and Retirement Study and single-cell RNAseq data. Based on joint work with Uriel Martínez León, Wanli Hong, and Yuliang Shi.
Date: Nov 12, 2026
Title: How the Spectrum Limits Convex Regularization
Abstract: How does the design matrix shape the limits of high-dimensional regression? Even when the sample size and total measurement strength are fixed, distributing that strength differently across directions can change estimation performance. In this talk, I will study this dependence for linear regression with right-rotationally invariant design matrices, a class that allows general singular-value distributions.
I will present a spectrum-dependent lower bound on the asymptotic mean-squared error of estimators using admissible coordinate-separable convex penalties. The bound comes from the state evolution of vector approximate message passing, which separates the effects of the signal distribution and the design spectrum through a scalar recursion. I will then characterize when a convex penalty can attain this bound: the criterion is log-concavity of the signal distribution after Gaussian smoothing at an effective noise level determined by the spectrum. Finally, I will explain how an ordering of spectral distributions yields comparisons between designs. In particular, mean-preserving spreads of the spectrum cannot decrease the lower bound, making precise how uneven measurement strength can limit estimation. Examples will illustrate both the spectral comparisons and the distinction between attainable bounds and strict gaps.
This is joint work with Baichen Tan and Audrey Yang.
Bio: Cynthia Rush is an Associate Professor of Statistics at Columbia University. Her research focuses on high-dimensional statistical inference and computationally efficient algorithms, drawing on information theory, statistical physics, and applied probability. She is particularly interested in approximate message passing and the statistical and computational limits of learning from data. She received her Ph.D. in Statistics from Yale University in 2016 and her B.S. in Mathematics from the University of North Carolina at Chapel Hill. She is a 2026 IEEE Information Theory Society Goldsmith Lecturer.
Date: Nov 19, 2026
Title:
Abstract:
Date: March 18, 2027
Title:
Abstract:
Date: March 25, 2027
Title:
Abstract:
Date: April 1, 2027
Title:
Abstract:
Date: April 8, 2027
Title:
Abstract:
Date: April 15, 2027
Title:
Abstract:
Date: April 22, 2027
Title:
Abstract: