# Forthcoming Seminars

Please note that the list below only shows forthcoming events, which may not include regular events that have not yet been entered for the forthcoming term. Please see the past events page for a list of all seminar series that the department has on offer.

## Further Information:

Speaker 1: Pawan Kumar

Title: Neural Network Verification

Abstract: In recent years, deep neural networks have started to find their way into safety critical application domains such as autonomous cars and personalised medicine. As the risk of an error in such applications is very high, a key step in the deployment of neural networks is their formal verification: proving that a network satisfies a desirable property, or providing a counter-example to show that it does not. In this talk, I will formulate neural network verification as an optimization problem, briefly present the existing branch-and-bound style algorithms to compute a globally optimal solution, and highlight the outstanding mathematical challenges that limit the size of problems we can currently solve.

Speaker 2: Samuel Albanie

Title: The Design of Deep Neural Network Architectures: Exploration in a High-Dimensional Search Space

Abstract: Deep Neural Networks now represent the dominant family of function approximators for tackling machine perception tasks. In this talk, I will discuss the challenges of working with the high-dimensional design space of these networks. I will describe several competing approaches that seek to fully automate the network design process and the open mathematical questions for this problem.

By viewing a stochastic process as a random variable taking values in a path space, the support of its law describes the set of all attainable paths. In this talk, we show that the support of the law of a solution to a path-dependent stochastic differential equation is given by the image of the Cameron-Martin space under the flow of mild solutions to path-dependent ordinary differential equations, constructed by means of the vertical derivative of the diffusion coefficient. This result is based on joint work with Rama Cont and extends the Stroock-Varadhan support theorem for diffusion processes to the path-dependent case.

Backward Stochastic Differential Equations (BSDEs) provide a systematic way to obtain Feynman-Kac formulas for linear as well as nonlinear partial differential equations (PDEs) of parabolic and elliptic type, and the numerical approximation of their solutions thus provide Monte-Carlo methods for PDEs. BSDEs are also used to describe the solution of path-dependent stochastic control problems, and they further arise in many areas of mathematical finance.

In this talk, I will discuss the numerical approximation of BSDEs when the nonlinear driver is not Lipschitz, but instead has polynomial growth and satisfies a monotonicity condition. The time-discretization is a crucial step, as it determines whether the full numerical scheme is stable or not. Unlike for Lipschitz driver, while the implicit Bouchard-Touzi-Zhang scheme is stable, the explicit one is not and explodes in general. I will then present a number of remedies that allow to recover a stable scheme, while benefiting from the reduced computational cost of an explicit scheme. I will also discuss the issue of numerical stability and the qualitative correctness which is enjoyed by both the implicit scheme and the modified explicit schemes. Finally, I will discuss the approximation of the expectations involved in the full numerical scheme, and their analysis when using a quasi-Monte Carlo method.

The concept of an acylindrically hyperbolic group, introduced by D. Osin, generalizes hyperbolic and relatively hyperbolic groups, and includes many other groups of interest: Out(F_n), n>1, most mapping class groups, directly indecomposable non-cyclic right angled Artin groups, most graph products, groups of deficiency at least 2, etc. Roughly speaking, a group G is acylindrically hyperbolic if there is a (possibly infinite) generating set X of G such that the Cayley graph \Gamma(G,X) is hyperbolic and the action of G on it is "sufficiently nice". Many global properties of hyperbolic/relatively hyperbolic groups have been also proved for acylindrically hyperbolic groups.

In the talk I will discuss a method which allows to construct a common acylindrically hyperbolic quotient for any countable family of countable acylindrically hyperbolic groups. This allows us to produce acylindrically hyperbolic groups with many unexpected properties.(The talk will be based on joint work with Denis Osin.)

TBA

With the introduction of supermarket loyalty cards in recent decades, there has been an ever-growing body of customer-level shopping data. A natural way to represent this data is with a bipartite network, in which customers are connected to products that they purchased. By predicting likely edges in these networks, one can provide personalised product recommendations to customers.

In this talk, I will first discuss a basic approach for recommendations, based on network community detection, that we have validated on a promotional campaign run by our industrial collaborators. I will then describe a multilayer network model that accounts for the fact that customers tend to buy the same grocery items repeatedly over time. By modelling such correlations explicitly, link-prediction accuracy improves considerably. This approach is also useful in other networks that exhibit significant edge correlations, such as social networks (in which people often have repeated interactions with other people), airline networks (in which popular routes are often served by more than one airline), and biological networks (in which, for example, proteins can interact in multiple ways).

In this talk, we analyze a virtual element method (VEM) for solving a non-selfadjoint fourth-order eigenvalue problem derived from the transmission eigenvalue problem. We write a variational formulation and propose a $C^1$-conforming discretization by means of the VEM. We use the classical approximation theory for compact non-selfadjoint operators to obtain optimal order error estimates for the eigenfunctions and a double order for the eigenvalues. Finally, we present some numerical experiments illustrating the behavior of the virtual scheme on different families of meshes.

When solving partial differential equations driven by additive spatial white noise, the efficient sampling of white noise realizations can be challenging. In this talk we focus on the efficient sampling of white noise using quasi-random points in a finite element method and multilevel Quasi Monte Carlo (MLQMC) setting. This work is an extension of previous research on white noise sampling for MLMC.

We express white noise as a wavelet series expansion that we divide in two parts. The first part is sampled using quasi-random points and contains a finite number of terms in order of decaying importance to ensure good QMC convergence. The second part is a correction term which is sampled using standard pseudo-random numbers.

We show how the sampling of both terms can be performed in linear time and memory complexity in the number of mesh cells via a supermesh construction. Furthermore, our technique can be used to enforce the MLQMC coupling even in the case of non-nested mesh hierarchies. We demonstrate the efficacy of our method with numerical experiments.