Gaussian processes for Science
Room: Kilburn Building, Lecture Theatre 1.1
8:30-9:00 Arrivals
9:00-9:10 Welcome
9:10-10:10 Modelling invariances and equivariances with GP models
10:10-10:40 Coffee Break
10:40-11:40 Emulating cohorts of cardiac digital twins using Gaussian Processes [slides]
11:40-12:40 Transfer Learning Gaussian Processes for DNA Design [slides]
12:40-14:00 Lunch
14:00-15:00 Using GP emulation in cardiovascular modelling [slides]
15:00-15:30 Tea Break
15:30-16:30 Applications of Gaussian Processes in Oceanography [slides]
Abstracts
Modelling invariances and equivariances with GP models
Abstract
Gaussian Process models offer elegant possibilities to encode invariances and equivariances. Choosing adapted mean and covariance functions enables going around data augmentation or making it somehow implicit. Furthermore, resulting models do not only propagate invariances and equivariances via the predictive mean but also via posterior simulations. We review related results and illustrations pertaining to invariances, and tackle recent work pertaining to equivariances, introducing in turn via examples some recent research about computationally efficient equivariant GP modelling. Based on several collaborations to be further detailed during the presentation.
Emulating cohorts of cardiac digital twins using Gaussian Processes
Abstract
Mathematical models of individual patients (digital twins) have the capacity to be a key tool for personalised medicine which will allow us to make data driven clinical recommendations in real time. However, sufficiently detailed cardiac models are often computationally expensive to run and this limits their use in fast-paced clinical settings. One method of decreasing these computational costs is to use Gaussian process emulators (GPEs) as surrogate models. However, training a GPE still requires a large number of simulations of the original model. In this talk we will discuss methods of cohort calibration, where new patient models learn from a cohort of existing personalised GPEs. These cohort approaches reduce the simulation overheads whilst maintaining similar predictive power to GPEs trained on large sets of simulation data.
Transfer Learning Gaussian Processes for DNA Design
Abstract
With the rise in engineered biomolecular devices, there is an increased need for tailor-made biological sequences. Often, many similar biological sequences need to be made for a specific application meaning numerous, sometimes prohibitively expensive, lab experiments are necessary for their optimisation. In this talk, I will present a transfer learning design of experiments workflow to make this development feasible. By combining a transfer learning surrogate Gaussian process model with Bayesian optimisation, I'll show how the total number of experiments can be reduced by sharing information between optimisation tasks. We will see how this method can be applied to data from the development of DNA competitors for use in an amplification-based diagnostic assay.
Using GP emulation in cardiovascular modelling
Abstract
There have been impressive advances in the physical and mathematical modelling of complex systems in the last few decades, and increasingly cover areas that until recently have been regarded as elusive for the quantitative sciences. The focus of this talk will be on cardiovascular modelling, which has the potential to revolutionise personalised healthcare with accurate patient-specific risk prediction and evidence-based treatment through digital replicas. The coupled non-linear partial differential equations that the cardiovascular models are based on depend on unknown physical parameters, as well as initial and boundary conditions, which are estimated from data. A fundamental challenge is the fact that the parameter inference and uncertainty quantification techniques require repeatedly getting outputs from the model for different parameter configurations typically thousands of times. This leads to very large computational times, making this approach practically infeasible in a clinical setting. I address this problem with emulation using Gaussian Processes (GPs), by developing a computationally tractable statistical surrogate model of the original intractable mathematical or physical model (the “digital twin”). In this talk I will present the application of GP emulation to two cardiovascular models: one of the pulmonary circulation, and a second of the coronary circulation.
Applications of Gaussian Processes in Oceanography
Abstract
Oceanography has long sought to understand the spatial and temporal structure of many different oceanographic processes: turbulent decay, tides, surface waves, internal waves, internal tides, and the list goes on. Most commonly, oceanographers use spectral analysis as the main tool to infer any parameters of interest, relying on regularly sampled observations of the particular process. This works well, for example, for mooring data that records temporal measurements at a single location in space, or for satellite data that records spatial measurements at single locations in time. Increasingly, we are building large data-sets that combine many different observation platforms with the aim to resolve the 3D (and 4D) spatio-temporal characteristics. Gaussian processes provide a principled method to merge all these data inside of a coherent inferential framework and are increasingly being recognised in oceanography. In this presentation, I will cover varied applications of Gaussian processes in oceanography, from relatively simple examples of temporal forecasting of surface tides, to some nascent attempts of inferring flow properties from multiple observation platforms. Finally, I’ll highlight some methodological gaps that we are looking to resolve.