Events

Talk description: I study how structure in graphs, such as decompositions and sparsity, can be uncovered and exploited to design efficient algorithms. In this talk, I will discuss how my work and vision connect foundational theory with practice through applications in database query optimization, pattern mining, and more.

Bio: Vaishali Surianarayanan is a Toni Massini Postdoctoral Fellow in Data Science at Yale University. She received her PhD from UC Santa Barbara and spent one year at UC Santa Cruz as a Chancellor’s Postdoctoral Fellow before joining Yale. Her research lies at the intersection of theoretical computer science, databases, and data mining.

A central theme of her work is using structural properties of data, such as sparsity, density, acyclicity, and graph decompositions, to design theoretically sound algorithms that can also inform practice. Much of her work is on graphs, with applications like database query evaluation, motif and subgraph counting, and relational clustering.

More broadly, she is interested in bridging foundational graph theory with real-world data systems. With this agenda in mind, she is especially interested in understanding structure in data across diverse domains, such as social networks, biological data, and knowledge graphs.

Her research has been recognized with the Mihai Pătrașcu Best Paper Award at SOSA 2026 and multiple “best of” journal invitations. She has also been invited to prestigious workshops including the Junior Theorists Workshop, UCR FAME, and Dagstuhl.

Website: https://vaishalisurianarayanan.weebly.com/

Talk description: Local SGD, also known as Federated Averaging, is a communication-efficient distributed optimization method in which each client performs several stochastic-gradient steps on its own data before communicating and averaging with the other clients. In this talk, I will ask when these sequential local updates outperform using the same stochastic gradients in a large mini-batch, and review a hierarchy of assumptions on data heterogeneity that helps answer this question. I will present upper and lower bounds showing when local updates help, with particular emphasis on how similarities in curvature across clients can permit substantial heterogeneity while retaining the benefits of Local SGD.

Speaker bio: Kumar Kshitij Patel joined Yale as a Postdoctoral Associate in the Summer of 2025. Kshitij received his PhD in Computer Science from the Toyota Technological Institute at Chicago, where he was advised by Prof. Nathan Srebro and Lingxiao Wang. Before that, he obtained his B.Tech. in Computer Science and Engineering from the Indian Institute of Technology (IIT) Kanpur. His research focuses on federated learning, optimization, and privacy, with a broader interest in how machine learning systems can efficiently and reliably operate across multiple tasks and data distributions under constrained access. His work combines theoretical frameworks, such as min-max complexity, with algorithm design to model and address key challenges arising from communication costs, data heterogeneity, fairness issues, and distribution shifts. He is particularly interested in domains such as healthcare, where both the need to integrate multiple datasets and the imperative to preserve privacy arise as central challenges. His research has been recognized with the Distinguished Paper Award at IJCAI 2024 and a Best Paper Honorable Mention at the Federated Learning Workshop at ICML 2023. As an FDS Postdoc, he is interested in the economics of data, specifically in how to assign value, design incentive mechanisms and markets, and protect privacy, fairness, and ownership as the demand for high-quality datasets increases. His goal is to help develop collaborative learning systems that are socially beneficial and practical, ensuring fair compensation for data owners, enabling trustworthy cooperation, and upholding data rights while informing legislation that governs data usage and enforces privacy protections.

Talk description: Across many physical and biological applications, one observes evolving probability distributions rather than individual particle trajectories. This raises a fundamental question: when can a dynamical system be uniquely identified and forecast from finitely many density snapshots? This talk establishes guarantees for the recovery of transport maps and vector fields from such measure-valued data, yielding new insights into generative models, data-driven dynamical systems, and PDE inverse problems.

Speaker bio: Jonah Botvinick-Greenhouse joined the Institute for Foundations of Data Science as a Postdoctoral Fellow in July 2026. He obtained his Ph.D. in Applied Mathematics from Cornell University, where he was advised by Professor Yunan Yang. During his Ph.D., he was an NDSEG Fellow and completed research internships at Mitsubishi Electric Research Laboratories and Argonne National Laboratory. Previously, he received B.A.s in Mathematics and Physics from Amherst College. His research interests lie at the intersection of data-driven dynamical systems, measure transport, scientific machine learning, numerical analysis, and inverse problems. In particular, his work explores connections between dynamical systems and measure transport, spanning theory, algorithms, and applications for learning and reconstructing complex systems from noisy, partially observed, or distributional data.

Website: jrbotvinick.github.io