Seminar: Intelligent Projector Systems for Spatial Augmented Reality
Haibin Ling, Stonybrook University
April 24th, 2025
12:45pm to 1:45pm (Free Period)
Armitage – 121
Abstract: The rapid advancement of imaging techniques and artificial intelligence has revolutionized research and applications in visual intelligence (VI). In this talk, I will present our studies covering a broad range of topics in VI, including visual recognition, video understanding, visual enhancement, and relevant machine learning techniques, with applications in virtual/augmented reality, biomedical research, and more.
I will then present our recent work applying AI to projector systems for spatial augmented reality tasks. In particular, image-based relighting, projector compensation and depth/normal reconstruction are three important tasks of projector-camera systems (ProCams) and spatial augmented reality (SAR). Although they share a similar pipeline of finding projector-camera image mappings, in tradition, they are addressed independently, sometimes with different prerequisites, devices and sampling images. In practice, this may be cumbersome for SAR applications to address them one-by-one. In this talk, we propose a novel end-to-end trainable model named DeProCams to explicitly learn the photometric and geometric mappings of ProCams, and once trained, DeProCams can be applied simultaneously to the three tasks. DeProCams explicitly decomposes the projector-camera image mappings into three subprocesses: shading attributes estimation, rough direct light estimation and photorealistic neural rendering. In our experiments, DeProCams shows clear advantages over previous arts with promising quality and meanwhile being fully differentiable. Moreover, by solving the three tasks in a unified model, DeProCams waives the need for additional optical devices, radiometric calibrations and structured light patterns. We will also briefly show our recent work on language-guided projection content generation. This is a joint work with Bingyao Huang.
Seminar: The AI’s impact on the Future of Education
Ananda Gunawardena, NJ AI Hub
April 22nd, 2025
12:45 PM to 1:45 PM (Free Period) in Armitage – 124 or join Online
followed by open discussion and refreshments at “The Alumni House” Cooper Street
This is a joint event hosted by the Research Office, Provost Office, and the Math department.
Abstract: As AI continues to transform education, we need to think about how AI will impact education and future workforce development. This seminar will highlight our work with NJ AI Hub (https://njaihub.org/), a centralized resource for foundational research, innovation economy, education and workforce development. We will also introduce key AI innovations, including Cubits.ai, an AI-driven course platform; CodeBench, a cloud-based coding environment; and SmartSlides, an AI-powered tool for creating dynamic presentations. These technologies aim to make learning more interactive, accessible, and effective. We will explore how AI is reshaping education and how new tools and course reimagining can support instructors and students in the evolving digital landscape
Seminar: Agent-Oriented Programming of Language Models
Wei Dong, National AI Campus and Managing Director of Ann Arbor Algorithms
April 17th, 2025
12:45 pm to 1:45 pm (Free Period)
Online
Abstract: A large language model is an automaton, and like any automaton, it should be programmed in the language it accepts. When we view prompt engineering as programming through the lens of automata theory, it becomes clear that traditional software engineering practice — rooted in the strict separation of programming languages and natural languages — must be rethought. In this seminar, Dr. Dong will present a conceptual framework of agent-oriented programming based on emails and share intriguing experimental results from Ann Arbor.
This is a joint event hosted by the Research Office, Provost Office, and the Math department.
Seminar: The Cauchy-Riemann Equations on Domains in the Complex Projective Space
Mei-Chi Shaw, University of Notre Dame
April 1st, 2025
11:00AM to 12:00PM
Armitage – 124
Abstract: The Cauchy-Riemann equations play central role in one and several complex variables. The Cauchy-Riemann operator ∂ has been studied extensively on domains in the complex Euclidean space Cⁿ. Much less is known when the ambient manifold is not Cⁿ.
In this talk, we discuss the range of ∂ on domains in the complex projective space CPⁿ. We also study the ∂-Cauchy problem on pseudoconvex domains and use it to prove the Sobolev estimates for ∂ on pseudoconcave domains in CPⁿ. In particular, we show that ∂ does not have closed range in L² for (2,1)-forms on the Hartogs triangle in CP². This is in sharp contrast to ∂ on the Hartogs triangle in C², where L² results have long been established by Hörmander.
Seminar: Continuity Equations in Fibered Wasserstein Spaces – A common framework for meanfield and graphon dynamics
Beniot Bonnet-Weill, Chargé de Recherche CNRS (Junior Researcher)
March 25th, 2025
12:45pm to 1:45pm (Free Period)
Armitage – 124
Abstract: During the past fifteen to twenty years, the the concept of meanfield approximation has become one of the leading paradigms in the mathematical analysis of large multiagent systems. This prominence can be explained by its mathematical versatility, its modelling power, and its general amenability to various families of numerical schemes.
However, meanfield limits are, by essence, confined to operating at the level of homogeneous particle systems, wherein the dynamics of each agent only depends on purely spatial quantities (e.g. its own position and that of the others). In order to provide a macroscopic description of heterogeneous multiagent systems, a more recent trend has consisted in leveraging the concept of graph limit, introduced by Lovasz and Szegedy. These are quite natural and relatively easy to manipulate, albeit a bit rigid as they lead to considering ODEs in Lebesgue spaces, coined graphon dynamics.
In this ongoing work in collaboration with Nastassia Pouradier Duteil (INRIA, Sorbonne Université), we investigate a new class of evolutions taking the form of continuity equations over spaces of Young measures endowed with an adequate “fibered” Wasserstein metric. The main interest in doing so is that the latter combine some of the desirable features of both meanfield and graphon dynamics, while providing an embedding of both in natural limit cases. In this context, I will present the basics of all three models, discuss some of the topological properties of fibered Wasserstein distances, and expose Carathéodory and Cauchy-Lipschitz well-posedness results for the underlying dynamics.
Seminar: Nonlocal Systems of Hyperbolic Equations
Mauro Garavello, University of Milano Bicocca
March 13th, 2025
12:45pm to 1:45pm (Free Period)
Armitage – 121
Abstract: We consider a multi non-linear system of hyperbolic equation in conservation form with non-local terms in the flux functions. The non-local terms considered here are convolutions with smooth kernels.
The resulting model is of macroscopic type and is able to describe different behaviors typically emerging in population dynamics. Indeed different shapes of the kernel function and of the velocity vectors may result from example in an aggregation phenomenon, with the possible formation of clusters (or opinions), or in a segregation of the various populations.
An important role is played by the support of the kernel function, which corresponds to the visual range of the individuals. In this talk we present several numerical integrations together with its well posedness and some analytic qualitative properties. Moreover, we also discuss the coupling between local and non-local equations. This interplay is of a particular interest in the context of traffic flow, due to the presence of autonomous vehicles, aware of traffic conditions at a significant distance from their locations, and of standard vehicles, which typically behaves according to the traffic conditions at their locations.
These are joint works with Rinaldo M. Colombo (University of Brescia, Brescia, Italy) and Claudia Nocita (University of Milano Bicocca, Milano, Italy).
Seminar: Why Neural Networks find simple solutions
Benoit Dherin, Google
March 10th, 2025
11:20am to 12:20pm (Free Period)
Armitage – 124
Abstract: Despite their ability to model very complicated functions and equipped with enough parameters to grossly overfit the training dataset, overparameterized neural networks seem instead to learn simpler functions that generalize well. In this talk, we present the notions Implicit Gradient Regularization (IGR) and Geometric Complexity (GC), which shed light on this perplexing phenomenon. IGR helps to guide the learning trajectory towards flatter regions in parameter space for any overparameterized differentiable model. This effect can be derived mathematically using Backward Error Analysis, a powerful and flexible method borrowed from the numerics of ODEs. For neural networks, we explain how IGR translates to a simplicity bias measured by the neural network GC. We will also show how various common training heuristics put a pressure on the GC, creating a built-in geometric Occam’s razor in deep learning.