Upcoming MAMS Seminar Series
Spring 2025
3/25/2025, Tue. 2:30-3:30 pm in Rockefeller 303
Speaker: Diliya Yalikun (Case Western Reserve University)
Title: Ball and spindle convexity with respect to a convex body
Abstract: Let C be a convex body, we introduce two notions of convexity relative to C. We will look at how some fundamental properties of classical convex sets can be applied to the C-ball convex sets and C-spindle convex sets.
3/18/2025, Tue. 2:30-3:30 pm in Rockefeller 303
Speaker: Effrosyni Chasioti (Case Western Reserve University)
Title: Dvoretszky’s Theorem
Abstract: We will state and present the proof of Dvoretszky’s Theorem. This theorem is one of the influential and fundamental results in convex geometry and the local theory of Banach spaces. It has been proved first by Dvoretszky in the 1961 and later was proven by V. Milman in 1971 using the concentration of measure phenomenon and connected this theorem with the local theory of Banach spaces in a more direct and revolutionary way.
This theorem basically says that every sufficiently high dimensional normed vector space has low dimensional subspaces that are approximately Euclidean, or in the language of convex geometry , that every high dimensional compact symmetric convex subset of $\mathbb{R}^N$ has a low dimensional sections that resemble ellipsoids.
3/4/2025, Tue. 2:30-3:30 pm in Rockefeller 303
Speaker: Mengchun Cai (Case Western Reserve University)
Title: Convergence Rate for some Classical Ensembles
Abstract: In this talk, I will introduce some classical ensembles in the random matrix theory and discuss their convergence rates to their limiting point processes with respect to the W_1 distance.
2/25/2025, Tue. 2:30-3:30 pm in Rockefeller 303
Speaker: Fabio Isaza (Case Western Reserve University)
Title: Approximating Continuous Functions with Neural Networks
Abstract: We will discuss the formal set up of neural networks as layered compositions of fixed non-linear and varying affine operators, the classical results by Hornik, Stinchcombe, and White; and more recent results by Boris Hanin and Mark Sellke connecting the NN functional approximations to approximation by free-node splines and more. We will also explore the theoretical underpinnings of Rectified Linear Unit (ReLU) NNs as universal approximators of continuous functions and discuss practical considerations in their implementation for function approximation tasks.
2/11/2025, Tue. 2:30-3:30 pm in Rockefeller 303
Speaker: Michael Roysdon (Case Western Reserve University)
Title: Comparison problems for Radon transforms
Abstract: In 1956 Herbert Busemann and Clinton M. Petty created a list of 10 problems in Convex Geometry, among which only the first had been solved fully in 1999. A key feature of the solution to the first problem is that it (and in fact, the majority of the Busemann-Petty problems) can be reformulated into the language of Harmonic Analysis. Inspired by the Busemann-Petty problems and their connection to Harmonic Analysis, we consider the following natural question for various Radon transforms: Let p>1. Given a pair of nonnegative, even and continuous functions f,g such that the Radon transform of f is pointwise smaller than the Radon transform of g, is it necessarily true that the L^p-norm of f is smaller than the L^p-norm of g? As it turns out, this simple question has a deep connection to the Busemann-Petty problem and the slicing problem of Bourgain. As a consequence of our investigation, we show that this implies reverse Oberlin-Stein type estimates for the spherical Radon transform when p >1; this is complementary to a recent work of Johnathan Bennett and Terence Tao in which similar reverse estimates were proven in the case 0 < p <= 1. We will discuss a similar problem for the dual Radon transform.
2/4/2025, Tue. 2:30-3:30 pm in Rockefeller 303
Speaker: Matthew Wascher (Case Western Reserve University)
Title: The effects of individual-level behavioral responses on SIS epidemic persistence
Abstract: The contact process (SIS epidemic model) has long been studied as a model for the spread of infectious disease through a population. One important question concerns the long-term behavior of the epidemic–does it result in a large outbreak, or does the infection die out quickly? There is a large literature on the effects of the underlying population structure on this long-term behavior. However, the role of individual-level behavioral responses on the epidemic is less studied. In this talk, I will introduce the contact process, some key ideas used to analyze it, and a few notable results. I will then discuss my recent work on modified versions of the contact process that include individual-level behavioral responses to the epidemic. I will present some results on how individual-level behavioral responses can influence the long-term behavior of an epidemic and discuss why analyzing these models is mathematically challenging.
Spring 2025
3/5/2025, Wed. 4-5 pm in Sears 439
Speaker: Yiming Ying (University of Sydney)
Title: Optimal Rates for Generalization of Gradient Descent Methods with ReLU Neural Networks
Abstract: The neural tangent kernel (NTK) perspective establishes a fundamental connection between neural networks (NNs) trained by gradient descent (GD) and kernel methods. A key question arising from this insight is whether GD-based training of NNs can achieve optimal statistical rates comparable to kernel methods.
In this talk, I will present our recent work addressing this question, offering new insights into the theoretical guarantees of gradient-based deep learning models. For binary classification, we show that gradient descent (GD) applied to training two-layer ReLU neural networks with the logistic loss can achieve the optimal margin bound, provided the data is NTK-separable. Our analysis highlights the intricate interplay between optimization and generalization, leveraging a reference model and a refined estimation of Rademacher complexity. For least-squares regression, we demonstrate that both GD and SGD with two-layer ReLU NNs can attain optimal minimax rates under polynomial overparameterization. Finally, I will discuss how these results extend to deep ReLU networks in the regression setting.
2/26/2025, Wed. 4-5 pm in Sears 439
Speaker: Anuj Abhishek (Case Western Reserve University)
Title: A probabilistic interpretation of Landweber iteration for a linear ill-posed inverse problem
Abstract: The contents of this talk are still at a very early proof-of-concepts stage. We will first describe the so-called Gaussian-beam Radon transform that is used to model the observed data in many optical tomography applications, such as optical coherence tomography (OCT), which use optical-beams instead of X-rays. X-rays can be modeled as traveling along straight lines, and as such, the model for the observed data in CT imaging is the Radon transform. However, monochromatic laser beams do not have a uniform shape intensity like X-rays. Since a Gaussian beam describes the characteristics of an electromagnetic beam, such as a laser: in many laser optics applications, the laser beam is assumed to be a Gaussian beam with a Gaussian intensity profile. It has been recently shown that in OCT, when a reconstruction algorithm does not account for the Gaussian beam (and instead the data is modeled by the usual Radon transform), the reconstruction quality greatly suffers. In this project, we model the observed data in OCT by a Gaussian-beam Radon transform and propose to use a Landweber type iterative method for reconstruction. In the second part of the talk, we present a probabilistic perspective on this iterative method. In this analysis, we will consider both the input (starting guess) as well as the output of the iterative method to be probability distributions. Thus each iteration of the method will be thought of as ‘pushforwarding’ a probability distribution. At the conclusion of the iterative method, we will use the mean of the output distribution as our estimate of the solution and use the output distribution to quantify the uncertainty in the estimate. We will also give some preliminary justification of the validity of the proposed approach. This is joint work with Souvik Roy (UT Arlington) and Anwesa Dey (University of Utah).
2/5/2025, Wed. 4-5 pm in Sears 439
Speaker: Professor Sakshi Arya (CWRU)
Title: Semi-Parametric Batched Global Multi-Armed Bandits with Covariates
Abstract: The multi-armed bandits (MAB) framework is a widely used approach for sequential decision-making, where a decision-maker selects an arm in each round with the goal of maximizing long-term rewards. Moreover, in many practical applications, such as personalized medicine and recommendation systems, feedback is delayed and provided in batches, contextual information is available at the time of decision-making, and rewards from different arms are related rather than independent.
We propose a novel semi-parametric framework for batched bandits with covariates and a shared parameter across arms, leveraging the single-index regression (SIR) model to capture relationships between arm rewards while balancing interpretability and flexibility. Our algorithm, Batched single-Index Dynamic binning and Successive arm elimination (BIDS), employs a batched successive arm elimination strategy with a dynamic binning mechanism guided by the single-index direction. We consider two settings: one where a pilot direction is available and another where the direction is estimated from data, deriving theoretical regret bounds for both cases. When a pilot direction is available with sufficient accuracy, our approach achieves minimax-optimal rates (with $d = 1$) for nonparametric batched bandits, circumventing the curse of dimensionality. Extensive experiments on simulated and real-world datasets demonstrate the effectiveness of our algorithm compared to the nonparametric batched bandit method.