Upcoming talks

Control variates for computing transport coefficients

Renato Spacek (École des Ponts ParisTech)

Monday 4 December 2023, 14:00-15:00
Room 104, Arts Building (R16)

In molecular dynamics, transport coefficients measure the sensitivity of the invariant probability measure of the stochastic dynamics at hand with respect to some perturbations. They are typically computed using either the linear response of nonequilibrium dynamics, or the Green-Kubo formula. The estimators for both approaches have extremely high variance, which motivates the study of variance reduction techniques for computing transport coefficients. We present an alternative approach, called the "transient subtraction technique" (inspired from early work by Ciccotti and Jaccucci in 1975), which amounts to simulating a transient dynamics, from which we subtract a sensibly coupled equilibrium trajectory, resulting in an estimator with smaller variance. In this talk, we present the mathematical formulation of the transient subtraction technique, give various error estimates on the bias and variance of the associated estimator, and demonstrate the relevance of the method through numerical illustrations for various systems.


Catherine Drysdale (University of Birmingham)

Monday 18 December 2023, 14:00-15:00
Room 104, Arts Building (R16)



Yi Yu (University of Warwick)

Thursday 1 February 2024, 14:00-15:00
Room 104, Arts Building (R16)



James Foster (University of Bath)

Thursday 8 February 2024, 14:00-15:00
Room 104, Arts Building (R16)


Recent talks

Nonlinear dynamics of recurrent neural network function and malfunction

Peter Ashwin (University of Exeter)

Monday 20 November 2023, 14:00-15:00
Room 104, Arts Building (R16)

Trained recurrent neural networks (RNNs) are nonlinear systems that can be trained to do a variety of tasks (for example, finite state computation) in the presence of input. In this talk I will discuss some work on how excitable network attractors can be used to explain some aspects of function and malfunction in such networks.

Smooth over-parametrized solvers for non-smooth structured optimisation

Clarice Poon (University of Warwick)

Monday 6 November 2023, 14:00-15:00
Room 104, Arts Building (R16)

Non-smooth optimization is a core ingredient of many imaging or machine learning pipelines. Non-smoothness encodes structural constraints on the solutions, such as sparsity, group sparsity, low-rank and sharp edges. It is also the basis for the definition of robust loss functions such as the square-root lasso. Standard approaches to deal with non-smoothness leverage either proximal splitting or coordinate descent. The effectiveness of their usage typically depend on proper parameter tuning, preconditioning or some sort of support pruning. In this work, we advocate and study a different route. By over-parameterization and marginalising on certain variables (Variable Projection), we show how many popular non-smooth structured problems can be written as smooth optimization problems. The result is that one can then take advantage of quasi-Newton solvers such as L-BFGS and this, in practice, can lead to substantial performance gains. Another interesting aspect of our proposed solver is its efficiency when handling imaging problems that arise from fine discretizations (unlike proximal methods such as ISTA whose convergence is known to have exponential dependency on dimension). On a theoretical level, one can connect gradient descent on our over-parameterized formulation with mirror descent with a varying Hessian metric. This observation can then be used to derive dimension free convergence bounds and explains the efficiency of our method in the fine-grids regime.

Bayesian estimation using loss functions

Yu Luo (King’s College London)

Monday 30 October 2023, 14:00-15:00
Room 104, Arts Building (R16)

In the usual Bayesian setting, a full probabilistic model is required to link the data and parameters, and the form of this model and the inference and prediction mechanisms are specified via de Finetti's representation. In general, such a formulation is not robust to model mis-specification of its component parts. An alternative approach is to draw inference based on loss functions, where the quantity of interest is defined as a minimizer of some expected loss, and to construct posterior distributions based on the loss-based formulation; this strategy underpins the construction of the Gibbs posterior. We develop a Bayesian non-parametric approach; specifically, we generalize the Bayesian bootstrap, and specify a Dirichlet process model for the distribution of the observables. We implement this using direct prior-to-posterior calculations, but also using predictive sampling. The two updating frameworks yield the same posterior distribution under the exchangeability assumption and guarantee consistent estimation under mild conditions. We also study the assessment of posterior validity for non-standard Bayesian calculations. The methodology is demonstrated via the semi-parameter linear model.

Parameter estimation for macroscopic pedestrian dynamics models using trajectory data

Susana Gomes (University of Warwick)

Monday 23 October 2023, 14:00-15:00
Room 104, Arts Building (R16)

In this talk I will present a framework for estimating parameters in macroscopic models for crowd dynamics using data from individual trajectories. I consider a model for the unidirectional flow of pedestrians in a corridor which consists of a coupling between a density dependent stochastic differential equation and a nonlinear partial differential equation for the density. In the stochastic differential equation for the trajectories, the velocity of a pedestrian decreases with the density according to the fundamental diagram. Although there is a general agreement on the basic shape of this dependence, its parametrization depends strongly on the measurement and averaging techniques used as well as the experimental setup considered. I will discuss identifiability of the parameters appearing in the fundamental diagram, introduce optimisation and Bayesian methods to perform the identification, and analyse the performance of the proposed methodology in various realistic situations. Finally, I discuss possible generalisations, including the effect of the form of the fundamental diagram and the use of experimental data.

Using supervised machine learning and evolutionary optimization to understand risk factors of dissociation in young people

Daniel Herring (University of Birmingham)

Monday 16 October 2023, 14:00-15:00
Room 104, Arts Building (R16)

Dissociation is a condition that disproportionately impacts young adults and impacts their daily lives; the mildest experiences are akin to ‘zoning out’ but more severe cases can include the feeling of detachment from reality, depersonalization (where your thoughts and feelings don’t belong to you) as well as identity confusion and amnesia. The causes and contributing factors for dissociation are not well understood and can vary, however the recently proposed ČEFSA-14 score allows the quantification of Felt Sense of Anomaly (FSA) subtype of dissociative experiences and thus for novel analyses. A dataset comprising 2384 responses from the general UK population including demographic information and psychological measurements including trauma, stress and anxiety, allows for analysis towards a multifactorial explanation of increased risk for dissociation. The application of a Naïve Bayes classifier (NBC) testing framework allows for the relative importance of different key predictors (measurements) to be ascertained. Evolutionary optimization is then used to determine the optimal subset from a wider range of predictors maximises the NBC classification accuracy and therefore highlights the most important factors influencing dissociation. Analysis on data partitioned by demographic information and the construction of high-risk profiles is undertaken to allow for meaningful clinical impact and actionable outcomes to arise from this analysis.

Recent Advances in Data Stream Learning

Shuo Wang (University of Birmingham)

Monday 9 October 2023, 14:00-15:00
Room 104, Arts Building (R16)

In many real-world applications, data are generated or collected continuously in the form of data streams, such as social media data analysis, spam detection and IoT systems. Data stream learning is such a machine learning area that processes data streams, trains a model and makes predictions over time. In this talk, four key research directions in data stream learning will be introduced with our recent findings and some applications, including class imbalance, concept drift, federated learning and AutoML.

Quantifying the economic and environmental effects of the RCEP

Kailan Tian (Chinese Academy of Sciences)

Monday 2 October 2023, 14:00-15:00
Room 104, Arts Building (R16)

Regional trade agreements (RTAs) have been widely adopted to facilitate international trade and cross-border investment and promote economic development. However, ex-ante measurements of the environmental effects of RTAs to date have not been well conducted. Here, we estimate the CO2 emissions burdens of the Regional Comprehensive Economic Partnership (RCEP) after evaluating its economic effects. We find that trade among RCEP member countries will increase significantly and economic output will expand with the reduction of regional tariffs. However, the results show that complete tariff elimination among RCEP members would increase the yearly global CO2 emissions from fuel combustion by about 3.1%, doubling the annual average growth rate of global CO2 emissions in the last decade. The emissions in some developing members will surge. In the longer run, the burdens can be lessened to some extent by the technological spillover effects of deeper trade liberalization. We stress that technological advancement and more effective climate policies are urgently required to avoid undermining international efforts to reduce global emissions.