Deep Learning for Model Calibration. EMPS College Home fees Studentship, PhD in Computer Science Ref: 4312
About the award
Prof Tim Dodwell - College of Engineering, Mathematics and Physical Sciences, Streatham Campus, University of Exeter
Dr George De Ath - College of Engineering, Mathematics and Physical Sciences, Streatham Campus, University of Exeter
Prof Richard Everson - College of Engineering, Mathematics and Physical Sciences, Streatham Campus, University of Exeter
Department of Computer Science, University of Exeter, Streatham Campus, Exeter, Devon.
The University of Exeter’s College of Engineering, Mathematics and Physical Sciences is inviting applications for a fully-funded PhD studentship to commence in January 2022 or as soon as possible thereafter. The studentship will cover Home tuition fees plus an annual tax-free stipend of at least £15,609 for 3.5 years full-time, or pro rata for part-time study.
This College studentship is open to UK and Irish nationals, who if successful in their application will receive a full studentship including payment of university tuition fees at the home fees rate.
Computer models, often known as simulators, are common in many scientific disciplines, such as climate science, epidemiology, engineering and public health. Frequently, these are expensive to run, and it is not uncommon for simulations to take days, or even weeks, to complete. The ability of the simulator to accurately model the physical or computational process of interest relies on its parameters being correctly calibrated. The task of model calibration is to automatically learn which parameters, or sets of parameters, allow the simulator to reproduce the observed process with minimal error.
Model calibration starts by building a surrogate model of the expensive simulator’s output, using previously-evaluated input parameters and their corresponding outputs. It then proceeds iteratively by first optimising a function of the surrogate model to select one (or more) sets of inputs that are informative to the calibration process, and then to evaluate them with the simulator. The surrogate model is then updated with the newly evaluated input/output pairs, and this process is repeated until a stopping criterion has been reached.
Gaussian processes (GP), the surrogate models of choice in calibration, are much less accurate with larger numbers of inputs. GPs are also usually limited to one output, meaning that as many GPs as simulator outputs are required unless the output has its dimensionality reduced by, for example, principal component analysis. Deep learning, i.e. neural network-based, approaches scale well with the number of both inputs and outputs, and are, therefore, a natural candidate to replace GPs in the calibration process. However, unlike GPs, deep learning-based methods do not inherently provide a level of uncertainty with their predictions.
Consequentially, this PhD will develop a deep learning-based framework for model calibration, with a particular focus on: creating deep learning surrogate models using less training data than traditional deep learning methods quantifying the amount of uncertainty in the surrogate model’s predictions and incorporating this into the calibration process investigating the use of both invertible neural networks and generative deep learning models to directly sample inputs that correspond to a outputs close to the target output.
The expected outcomes of the proposed project includes papers submitted to top machine learning conferences and the development of open source software as a result of the work carried out in the PhD to enable general-purpose high dimensional model calibration.
This studentship is open to UK and Irish nationals, who if successful in their application will receive a full studentship including payment of university tuition fees at the home fees rate.
Applicants for this studentship must have obtained, or be about to obtain, a First or Upper Second Class UK Honours degree, or the equivalent qualifications gained outside the UK, in an appropriate area of science or technology.
Candidates should have a strong mathematical background, specifically in probability and statistics. Experience of coding in Python and the use of deep learning libraries, such as PyTorch or Tensorflow, is desirable.
If English is not your first language you will need to have achieved at least 6.0 in IELTS and no less than 6.0 in any section by the start of the project.
Alternative tests may be acceptable (see http://www.exeter.ac.uk/postgraduate/apply/english/).
How to apply
In the application process you will be asked to upload several documents. Please note our preferred format is PDF, each file named with your surname and the name of the document, eg. “Smith – CV.pdf”, “Smith – Cover Letter.pdf”, “Smith – Transcript.pdf”.
• Letter of application (outlining your academic interests, prior project work experience and reasons for
wishing to undertake the project).
• Transcript(s) giving full details of subjects studied and grades/marks obtained (this should be an
interim transcript if you are still studying)
• Two references from referees familiar with your academic work. If your referees prefer, they can email
the reference direct to firstname.lastname@example.org quoting the studentship reference number.
• If you are not a national of a majority English-speaking country you will need to submit evidence of your proficiency in English. Please see the entry requirements information above.
The closing date for applications is midnight on 10th January 2022. Interviews will be held online on the week commencing 24th January 2022
If you have any general enquiries about the application process please email email@example.com.
Project-specific queries should be directed to the main supervisor at T.Dodwell@exeter.ac.uk
|Application deadline:||10th January 2022|
|Value:||Home tuition fees plus an annual tax-free stipend of at least £15,609 for 3.5 years full-time, or pro rata for part-time study.|
|Duration of award:||per year|
|Contact: PGR Admissions Officefirstname.lastname@example.org|