Time: 4:30-5:30 p.m.
Date: Wednesday, November 11, 2026
Location: 350 Computing and Information Science Building
Speaker: Subhro Ghosh, National University of Singapore
Title: Strongly correlated particle systems: Tools for kernelized sampling
Abstract: Independent and identically distributed (iid) random variables form a classical paradigm for randomness in the sciences, and departures from independence are often viewed as a challenge to overcome. In this talk, we explore a different perspective: strongly interacting particle systems from statistical and quantum physics can offer powerful tools for addressing fundamental problems in machine learning and statistics. Our focus will be on determinantal point processes (DPPs).
We will leverage these models for the fundamental task of sampling, with applications to stochastic gradient descent (SGD). In particular, we will show how computationally feasible kernelized samplers can achieve guarantees beyond those of iid sampling, in fact attaining information-theoretically optimal rates. Along the way, we will draw on ideas and tools from across applied mathematics and statistics, including orthogonal polynomials, wavelets, and spectral and diffusion geometry on graphs and manifolds. The talk highlights how advances in the theory of these stochastic models and in their applications can inform and enrich one another, opening up new directions for both.
Based in part on the following works:
[1] Gaussian determinantal processes: A new model for directionality in data,
(with P. Rigollet)
[Proceedings of the National Academy of Sciences (PNAS), vol. 117, no. 24, pp. 13207--13213]
[2] State-of-art minibatches via novel DPP kernels: discretization, wavelets, and rough objectives,
(with HS Tran, P Gupta, R Bardenet)
[NeurIPS 2026]
[3] Small coresets via negative dependence: DPPs, linear statistics, and concentration,
(with R. Bardenet, H. Simon-Onfroy, H.S. Tran)
[Spotlight at NeurIPS 2024]
[4] Determinantal point processes based on orthogonal polynomials for sampling minibatches in SGD,
(with R Bardenet, M Lin)
[Spotlight at NeurIPS 2021]
[5] Determinantal sampling on general spaces: fast approximation via diffusion geometry
(with HS Tran, P Gupta)
Bio: Subhro Ghosh is Dean's Chair Associate Professor at the National University of Singapore in the Dept of Mathematics, with joint appointments in Statistics & Data Science and Computer Science. He is also a faculty affiliate at the NUS AI Institute. Subhro is broadly interested in stochastics, focussing on problems from the math of machine intelligence, statistical physics and their interactions. Subhro obtained his PhD from UC Berkeley, and was a postdoc at Princeton University. His interests encompass a broad swathe of problems that concern learning under complex structure (e.g. latent symmetries, equivariances or community patterns), and interface with fundamental problems such as dimension reduction, sampling and optimization, statistical network models and signal processing. The overarching goal of his research is to develop a coherent understanding of how modern machine intelligence models work, and to leverage this understanding to the end of resource efficient and explainable AI systems.