Research seminars

Start at 15:00 via Zoom: https://bruneluniversity.zoom.us/j/95070805299
- Meeting ID: 950 7080 5299 (use Passcode as communicated per email)

ABSTRACT: To be updated.

Start at 15:00 via Zoom: https://bruneluniversity.zoom.us/j/95070805299
- Meeting ID: 950 7080 5299 (use Passcode as communicated per email)

ABSTRACT: To be updated.

Start at 15:00 via Zoom: https://bruneluniversity.zoom.us/j/95070805299
- Meeting ID: 950 7080 5299 (use Passcode as communicated per email)

ABSTRACT: To be updated.

Start at 15:00 via Zoom: https://bruneluniversity.zoom.us/j/95070805299
- Meeting ID: 950 7080 5299 (use Passcode as communicated per email)

ABSTRACT: To be updated.

Start at 15:00 via Zoom: https://bruneluniversity.zoom.us/j/95070805299
- Meeting ID: 950 7080 5299 (use Passcode as communicated per email)

ABSTRACT: To be updated.

Start at 15:00 via Zoom: https://bruneluniversity.zoom.us/j/95070805299
- Meeting ID: 950 7080 5299 (use Passcode as communicated per email)

ABSTRACT: We investigate the issue of parameter estimation with nonuniform negative sampling for imbalanced data. We first prove that, with imbalanced data, the available information about unknown parameters is only tied to the relatively small number of positive instances, which justifies the usage of negative sampling. However, if the negative instances are subsampled to the same level of the positive cases, there is information loss. To maintain more information, we derive the asymptotic distribution of a general inverse probability weighted (IPW) estimator and obtain the optimal sampling probability that minimizes its variance. To further improve the estimation efficiency over the IPW method, we propose a likelihood-based estimator by correcting log odds for the sampled data and prove that the improved estimator has the smallest asymptotic variance among a large class of estimators. It is also more robust to pilot misspecification. We validate our approach on simulated data as well as a real click-through rate dataset with more than 0.3 trillion instances, collected over a period of a month. Both theoretical and empirical results demonstrate the effectiveness of our method.

Start at 13:30 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: Two short talks on inference for time-varying networks.
1) Distinguishing mixtures of network growth models that vary in time
2) Studying an alt-right social network through the lens of temporal graphs

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: Statistically designed experiments are the “gold standard” for learning about products, processes and systems through the collection of data. By deliberately introducing controlled variability, whilst working to minimise uncontrolled variation, we can establish causal relationships, screen for important variables and build predictive models. I will describe how design of experiments and statistical modelling can go beyond the usual factorial design and response surface methodology. In particular, I will present methodology for (i) experiments with dynamic input variables; (ii) design to learn unknown parameters in empirical and first-principle nonlinear models; and (iii) Bayesian nonparametric learning through sequential experimentation. Where possible, methods will be illustrated on relevant examples.

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758  
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: Parametric partial differential equations (PDEs) occur in optimisation problems and in mathematical models with inherent uncertainties (e.g., groundwater flow models). Adaptive algorithms are indispensable when solving a particularly challenging class of parametric problems represented by PDEs whose inputs depend on infinitely many uncertain parameters. For this class of problems, adaptive algorithms have been shown to yield approximations that are immune to the curse of dimensionality -- an exponential growth of the computational cost as the dimension of the parameter space increases. In particular, adaptivity is the key to efficient stochastic Galerkin finite element methods (SGFEMs), where approximations are represented as finite (sparse) generalised polynomial chaos expansions with spatial coefficients residing in finite element spaces. While in the simplest (so-called single-level) SGFEM all spatial coefficients reside in the same finite element space, a more flexible multilevel construction allows spatial coefficients to reside in different finite element spaces.
     In this talk, we first give an overview of existing adaptive SGFEM algorithms and the associated theoretical results. Then, we focus on a recently proposed adaptive algorithm for computing multilevel stochastic Galerkin finite element approximations. We will discuss the convergence and rate optimality properties of the proposed algorithm and demonstrate its performance in numerical experiments.

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: Persistent homology is an algebraic tool for quantifying topological features of shapes and functions, which has recently found wide applications in data and shape analysis. In the first part of this talk, I aim to present the underlying algebraic ideas and basic concepts of this very active field of research. In the second part I sketch an application in digital image analysis.

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: The professional networking website LinkedIn and McKinsey Global Institute have listed skills in Machine Learning, AI, Statistics and Data Science as the top job demand. And FORTUNE (https://fortune.com/) identified that Statistics degrees lead careers in its Best and Worst Graduate Degrees for Jobs. I regularly receive queries about our MSc and PhD in SDS, from students/people who want to apply for courses here. I also recently talked to TNE students. They include "as the courses are in a Maths dept., how much Maths do I need?"; "My background is not in Maths but did few courses on it, can we do a good degree here?". My answer to them is that they need a university Maths year 1 and 2 "rapid" (regression + algebra + probability + integration + differentiation). Basic 'rapid' can apply for our SDA courses, but good 'rapid' may achieve a lot. This talk aims to demonstrate some of examples that the first/second year of UG Maths "rapid" do help me to develop new methods for modern statistical modelling and data analysis with good research impact.

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/98613510007
- Meeting ID: 986 1351 0007 (use Passcode as communicated per email)

ABSTRACT:In this talk we present three topics that we tackle using methods developed in the collective insurance risk literature. Specifically, we discuss insurance mechanisms for 1) mortgage lending, 2) car driving behaviour, and 3) poverty reduction.

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: Many studies collect multiple omics datasets to gather novel insights about different stages of biological processes. For joint modelling of these datasets, several data integration methods have been developed. These methods address high dimensionality of the datasets, within and across datasets correlation, and the presence of heterogeneity among datasets due to measuring different biological levels and using different technologies to measure them. Most methods, however, neither provide statistical evidence for a relationship between the datasets nor identify relevant variables that contribute to this relationship. We propose a probabilistic latent variable modelling framework for inferring the relationship between two omics datasets. Latent variable methods reduce dimensionality and capture correlations by forming components that are linear combinations of the variables. The correlation structure is modelled by joint and data specific components. We propose maximum likelihood estimation of the parameters and formulate a test statistic for the null hypothesis of no relationship between the datasets. We evaluate our methods via a simulation. Under the null hypothesis, the test statistic appears to approximately follow the normal distribution. Our method outperforms existing methods for small and heterogeneous datasets in terms of selecting relevant variables and prediction accuracy. We illustrate the methods by analysing omics datasets from a population cohort and a case control study. The obtained results are reproducible and biologically interpretable.

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: In the context of a Gaussian multiple regression model, we address the problem of variable selection when in the list of potential predictors there are factors, i.e., categorical variables. We adopt a model selection perspective, that is, we approach the problem by constructing a class of models, each corresponding to a particular selection of active variables. The methodology is Bayesian and proceeds by computing the posterior probability of each of these models. We highlight the fact that the set of competing models depends on the dummy variable representation of the factors, an issue already documented by Fernandez et al. (2002) in a particular example but that has not received any attention since then. We construct methodology that circumvents this problem and that presents very competitive frequentist behavior when compared with recently proposed techniques. Additionally, it is fully automatic, in that it does not require the specification of any tunning parameters. [Joint work with Gonzalo Garcia-Donato.]

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: There is no doubt that humans and other animals can not only recognise shapes but also differentiate shape from form. The latter is what aids us in classifying objects by eye, which is usually done in conjunction with more detailed descriptions of the object, such as its texture or patterns within. Subsequently, this leads to two questions which carve the basis of our work: (i) can an algorithm classify shapes in the same way a human does, and (ii) is the shape of an object alone sufficient for performing analysis and classification? To answer these questions, we will investigate and compare various non-linear and linear methods that quantify differences between shapes on a series of real-world datasets before moving to machine learning to perform classification. All the while, we will be working hand in hand with specialists in these fields, ranging from botanists to pottery historians. We are interested in analysing the shape of an object; crucially this means that we will not be studying landmarks as is often done with real-world data. Instead, we define the shapes as closed curves in R2 and work with the shape space of these curves using methods that are seldom applied to real-world data. This work is done in collaboration with Stephen Marsland (Victoria University of Wellington) and Armand Leroi (Imperial College London).

Start at 16:00 via Zoom: https://bruneluniversity.zoom.us/j/94345153758
- Meeting ID: 943 4515 3758 (use Passcode as communicated per email)

ABSTRACT: In causal inference for observational studies, we lack a randomised control for a treatment of interest. However, we sometimes have access to an instrumental variable, which is a type of (pseudo) randomisation that doesn’t fully control the treatment. In general, the causal effect still remains unidentifiable, but it may be possible to bound it. Until recently, no methods existed for the case where the treatment is a continuous variable. We will present an algorithm that facilitates this type of inference, while allowing for a continuum of causal assumptions that trade-off constraints on the unmeasured confounding structure against the informativeness of the bounds. Joint work with Niki Kilbertus (Helmholtz AI) and Matt Kusner (UCL).

Start at 15:30 via Zoom (joining instructions follow below)

ABSTRACT: Kernel based methods are widely used in both machine learning and Bayesian statistics. Standard kernels, like Gaussian and Matern kernels, are both stationary and separable. We introduce a flexible method to create nonstationary and non-separable kernels based on an infinite mixture of convolved stochastic processes. When the mixing process is stationary but the convolution function is nonstationary we arrive at non-separable kernels with constant non-separability that are available in closed form. When the mixing is nonstationary and the convolution function is stationary we arrive at non-separable random fields that have varying non-separability and better preserve local structure. We show how a single Gaussian process (GP) with these random fields can computationally and statistically outperform both separable and existing nonstationary non-separable approaches such as treed GPs and deep GP constructions.

Zoom link: https://bruneluniversity.zoom.us/j/99894030917
- Meeting ID: 998 9403 0917
- Passcode: 2108619563

Start at 15:30 via Zoom (joining instructions follow below).

ABSTRACT: In quantum mechanics, the direction of the probability density flow is not necessarily the same as that of velocity. One manifestation of this counter-intuitive fact is the so-called quantum backflow effect, in which the probability of finding a particle "on the left" increases with time despite the particle's velocity pointing "to the right." In my talk, I will provide a brief review of the quantum backflow problem, and will show how it can be generalized to quantum states with position-momentum correlations.
References: A. Goussev, arXiv:2002.03364 (2020); A. Goussev, PRA 99, 043626 (2019)

Zoom link: https://bruneluniversity.zoom.us/j/94925713575
- Meeting ID: 949 2571 3575
- Password: 8988773309

Start at 15:30 via Zoom (joining instructions follow below).

ABSTRACT: In quantum mechanics, the direction of the probability density flow is not necessarily the same as that of velocity. One manifestation of this counter-intuitive fact is the so-called quantum backflow effect, in which the probability of finding a particle "on the left" increases with time despite the particle's velocity pointing "to the right." In my talk, I will provide a brief review of the quantum backflow problem, and will show how it can be generalized to quantum states with position-momentum correlations.
References: A. Goussev, arXiv:2002.03364 (2020); A. Goussev, PRA 99, 043626 (2019)

Zoom link: https://bruneluniversity.zoom.us/j/94925713575
- Meeting ID: 949 2571 3575
- Password: 8988773309

Start at 15:30 via Zoom (joining instructions follow below).

ABSCTRACT: In this talk, I will propose a continuous-time stochastic intensity model, namely the two-phase dynamic contagion process, for modelling the epidemic contagion of COVID-19. The dynamic contagion process is a branching process which is an extension of the classical Hawkes process commonly used for modelling earthquakes and more recently financial contagion. In contrast to most existing models, this model allows randomness to the infectivity of individuals rather than a constant reproductive rate as assumed by standard models. Key epidemiological quantities such as the distribution of final epidemic size and expected epidemic duration have been derived and estimated based on real data for various regions and countries. The time that governmental intervention became effective is estimated and the results are consistent with the actual time of intervention and incubation time of the disease. The purpose of this talk will be to introduce the model and to show that this model could potentially be a valuable tool in the modelling of COVID-19 infection. Similar variations of this model could also be used in more general epidemiological modelling.

Join Zoom Meeting: https://bruneluniversity.zoom.us/j/98209870169
- Meeting ID: 982 0987 0169
- Password: 6659222909

Start at 15:30 via Zoom (joining instructions follow below).

ABSTRACT: Linear wave scattering problems (e.g. for acoustic, electromagnetic and elastic waves) are ubiquitous in science and engineering applications. However, conventional numerical methods for such problems (e.g. FEM or BEM with piecewise polynomial basis functions) are prohibitively expensive when the wavelength of the scattered wave is small compared to typical lengthscales of the scatterer (the so-called "high frequency" regime). This is because the solution possesses rapid oscillations which are expensive to capture using conventional approximation spaces. In this talk we outline recent progress in the development of "hybrid numerical-asymptotic" methods. These methods use approximation spaces containing oscillatory basis functions, carefully chosen to capture the high frequency asymptotic behaviour, leading to a significant reduction in computational cost.

Join Zoom Meeting: https://bruneluniversity.zoom.us/j/92367721405
- Meeting ID: 923 6772 1405
- Password: 7129906480

Start at 15:30 via Zoom (joining instructions follow below).

ABSTRACT: This presentation is split into three parts. The first part is a brief tutorial on how a commodity futures market operates, and the role played by stochastics in pricing of financial derivatives, including commodity futures. In the second part, I will introduce a recursive conditional mean estimator for linear Gaussian dynamic systems, popularly called the Kalman filter (KF), and its application in modelling the movement of crude oil futures prices.  Finally, I will describe how a KF-based model can be modified using quantitative measures of macroeconomic news sentiment to enhance the futures price prediction. I will also present a summary of our findings from numerical experiments on modelling and predicting crude oil futures prices using the Kalman filter and news sentiment scores.

Join Zoom Meeting at https://bruneluniversity.zoom.us/j/98340190032
- Meeting ID: 983 4019 0032
- Password: 0154229543

Start at 16:00 via Teams.

ABSCTRACT: I will be discussing a new methodology that permits the Bayesian supervised learning of generic tensor-valued functional relationships between two random variables, at least one of which is high-dimensional and distributed discontinuously. The ultimate aim is the prediction of either variable, at test data on the other. We will model the sought function as a random realisation from a tensor-variate, non-stationary stochastic process, each parameter of the covariance structure of which, is made to adapt to the discontinuity of the sought function. This is proven to be permitted by compounding multiple scalar-valued stationary processes with the mother tensor-valued non-stationary process. That 2 layers suffice in this learning strategy, is proven in the method. The talk will be motivation-heavy.

Start at 13:00 in TOWA-203

ABSTRACT: Random matrix theory has proven very successful in the understanding of the spectra of chaotic systems. Depending on symmetry with respect to time reversal and the presence or absence of a spin 1/2 there are three ensembles, the Gaussian orthogonal (GOE), Gaussian unitary (GUE), and Gaussian symplectic (GSE) one. With a further particle-antiparticle symmetry the chiral variants of these ensembles, the chiral orthogonal, unitary, and symplectic ensembles (the BDI, AIII, and CII in Cartan's notation) appear. A microwave study of the chiral ensembles is presented using a linear chain of evanescently coupled dielectric cylindrical resonators [1]. In all cases the predicted repulsion behavior between positive and negative eigenvalues for energies close to zero could be verified.
[1] https://arxiv.org/abs/1909.12886

Start at 11:00 in TOWA-203

ABSTRACT: Experimental design is the area of statistics which attempts to gain the most information out of an experiment or study, with minimal experimental effort. This talk provides a general introduction to a particular problem in this field for a general mathematical audience. We address the problem of how to optimally design experiments for subjects who are connected by a network structure. We also argue that there is a wide class of experiments that can be reformulated into a problem of design on a network. By regarding experimental design as a problem in network science, we can improve experimental design algorithms for large networks, and also find designs even when there is no obvious network relationship.