|
|
|
Invited speakersRT-UQ PhD day keynote speaker Jessica Hoffmann (Google) SAMO keynote speakers Tim Bedford (University of Strathclyde, Glasgow) Sergei Kucherenko (Imperial College London) Frances Kuo (University of New South Wales, Sydney) Francesca Pianosi (University of Bristol) Arnald Puy (University of Birmingham)
Keynote titles/abstracts RT-UQ PhD day Title: Epidemics on Graphs under Uncertainty Epidemic processes can model anything that spreads. As such, they are a useful tool for studying not only human diseases, but also network attacks, spikes in the brain, the propagation of real or fake news, the spread of viral tweets, and other processes. This talk focuses on epidemics spreading on an underlying graph. Currently, most state-of-the-art research in this field assumes some form of perfect observation of the epidemic process. This is an unrealistic assumption for many real-life applications, as the recent COVID-19 pandemic tragically demonstrated: data is scarce, delayed, and/or imprecise for human epidemics, and symptoms may appear in a non-deterministic fashion - if they appear at all. We show in this work not only that the algorithms developed previously are not robust to adding noise into the observation, but that some theoretical results cannot be adapted to this setting. In other words, uncertainty fundamentally changes how we must approach epidemics on graphs.
SAMO conference Title: Expertise, Uncertainty, and Modelling Recent calls for "Responsible modelling" (for example in Saltelli, A., G. Bammer, et al (2020). "Five ways to ensure that models serve society: a manifesto." Nature 582(7813): 482-484; and, Saltelli, A. and M. Di Fiore (2023). The politics of modelling: Numbers between science and policy.) from people with wide expertise in modelling and policy making, have set out challenges to the modelling community that require a conceptual response alongside a purely technical one. While the challenges they set out are wide ranging, we collectively need to make a start to constructively address some of these issues. Here, we argue that conventional uncertainty frameworks offer a mechanism for communication, between scientists about the possible forms of modelling and the state of the art, and with policy makers about the state of the (currently) possible, but do not provide an operational and formative approach for problem owners. They do not give an adequate framework to include the different types of knowledge available or take account of Box's famous view that "All models are wrong but some are useful". Here we develop a different approach which acknowledges the many roles of expertise alongside the possibilities of using multiple models of different types. These contribute to the form and content of modelling and hence impact on the reproducibility of the modelling process.
Title: Derivative Based Global Sensitivity Measures: Past, Present and Future
Title: Lattice rules, kernel methods, DNNs, and how to connect them Lattice rules are my favorite family of quasi-Monte Carlo methods. They are proven to be effective for high dimensional integration and multivariate function approximation in a number of settings. They are extremely easy to implement thanks to their very simple formulation --- all we require is a ``good'' integer vector of length matching the dimensionality of the problem. We know how to construct such good vectors tailored to applications in different areas, e.g., in PDEs with random coefficients, both for computing expected values (integrals) of quantities of interest as well as in obtaining surrogates of the PDE solution using lattice-based kernel interpolants. In recent years there has been a burst of research activities on the application and theory of Deep Neural Networks (DNNs). We explore how lattice rules can be used in the framework of DNNs. This is based on joint work with Alexander Keller (NVIDIA), Dirk Nuyens (KU Leuven) and Ian H. Sloan (UNSW Sydney).
Title: Global Sensitivity Analysis: who, when and why Global Sensitivity Analysis is increasingly used to investigate the propagation of uncertainties through environmental and infrastructure models. Knowledge of model outputs’ sensitivity can be used to guide the model calibration and diagnostic evaluation, and the use of model outputs for informing decisions under uncertainty. However, despite significant advances in the availability of methods, tools and application examples, the uptake of GSA widely varies within and across modelling communities. In this talk, I will discuss two “GSA paradoxes”. First, that the very complex models that would most benefit from scrutiny through GSA are the ones to which this methodology is least frequently applied. Second, that the versatility of GSA – i.e. its ability to be adapted to different tasks, from prioritising efforts for model improvement, to model evaluation, to improving our understanding of systems’ behaviour – makes it more difficult (rather than less) to communicate its value to potential users. Drawing on a range of recent examples from the water, natural risk and energy sector, I will present some ideas on how we can move forward past these paradoxes and better support model developers and users to conceptualise GSA experiments, and identify who, when and why can benefit from it.
Title: Smoke and mirrors in water modelling In this keynote I will survey our recent work on uncertainties in water modelling. I will show that knowledge claims in water modelling are as assertive and even more quantified than those in physics-based disciplines, yet their numeric inferences lack an uncertainty and sensitivity analysis (UA/SA). I will show what happens when one of the most spread claims, that humans have exceeded the freshwater planetary boundaries, is examined through a stringent UA/SA. Finally, I will zoom into irrigation modelling to highlight some consequences derived from this cursory approach to uncertainties and sensitivities: delusive accuracy in global irrigation water withdrawal estimates, excess of model complexity given epistemic and empirical limitations, and poor reflective stance, leading to a purely technical treatment of uncertainties. I will conclude by discussing why we should dispel the accuracy conjuring trick in water modelling and by offering some possible ways forward.
|
| Online user: 3 | Privacy |
|
