Skip to main content
School of Physical and Chemical Sciences

Breadcrumb

  1. School of Physical and Chemical Sciences
  2. Centre for Condensed Matter Physics
  3. PhD at CCMP
  4. PhD Projects in CCMP
  5. Developing Bayesian neural networks for applications in the transport, energy and environment sector

Developing Bayesian neural networks for applications in the transport, energy and environment sector

Research Group: Center for Condensed Matter Physics
Number of Students: 1
Length of Study in Years: 4 Years
Full-time Project: yes

Funding

This project has been supported by the Faculty for Chinese Scholarship Council (CSC) funding. If you wish to be considered for another funding route, please contact the supervisor [a.j.drew@qmul.ac.uk].

Project Description

Modern deep learning methods, in particular neural networks (NNs), are powerful tools to tackle a myriad of challenging problems, for example pattern recognition and classification across the domains of computer vision and speech recognition. However, deep learning methods operate as “black boxes”, with the uncertainty associated with their predictions being challenging to quantify. Increasing the number of parameters or the depth of the model increases the capacity of the network, allowing it to represent functions with greater non-linearities. Whilst this increase in capacity allows for more complex problems to be addressed, it leaves NNs highly prone to overfitting to the training data.

Since the deployment of NNs in the real world, there have been a number of incidents where their failings have led to models acting unethically and unsafely. This includes considerable sex bias and racism, and in more extreme cases, death. This lack of interpretation and over-confident estimates provided by the commonly used NNs makes them unsuitable for high risk domains such as medical diagnostics, autonomous vehicles and energy security. Bayesian Neural Networks offer natural ways to estimate uncertainty in predictions, and can provide insight into how decisions are made. However, calculating the exact Bayes factor is historically computationally expensive, and not accessible for embedded devices or desktop PCs.

We have recently demonstrated a method to calculate the Bayes factor exactly for a very low computational cost, such that it adds a negligible additional time to a standard least squares fitting routine [1]. This PhD will adapt these methods in a NN, so that exact calculations of the Bayes factor can be included routinely, thus enabling trustworthy models that can identify uncertainty in a prediction.

Please note this fits within two Faculty themes: Green Energy and AI/Data Modelling.

Supervisor Contact Details:

For informal enquiries about this position, please contact Alan Drew

E-mail: A.J.Drew@qmul.ac.uk

Deadline - 31st of January 2023

Application Method:

To apply for this studentship select September entry in the following page:

https://www.qmul.ac.uk/postgraduate/research/subjects/physics.html

[1] Dunstan, Crowne and Drew, Scientific Reports 12, 993 (2022)

Requirements

  • CSC, self-funded, external scholarship, CONACyT
  • Applicant required to start in September/October 2023.
  • The minimum requirement for this studentship opportunity is a good Honours degree (minimum 2(i) honours or equivalent) or MSc/MRes in a relevant discipline.
  • If English is not your first language you will require a valid English certificate equivalent to IELTS 6.5+ overall with a minimum score of 6.0 in Writing and 5.5 in all sections (Reading, Listening, Speaking).

SPCS Academics: Prof. Alan Drew

Back to top