Overview
Overview*
This is a short summary that should capture the overall aim of the project, highlight USPs, and inspire the reader’s passion for the project. 20-60 words recommended.
Transforming Medical AI: Accelerating Bayesian Neural Networks to provide interpretable, uncertainty-aware healthcare solutions with faster, energy-efficient computational methods from Bayesian Inference.
About this opportunity
Deep learning and artificial intelligence (AI) have promised transformative impact in the health sciences: from automatic detection of tumours in medical imaging, to natural language processing of electronic health records. These methods, however, face significant challenges in clinical uptake, partly due to a lack of interpretability around uncertainty of their outputs. Bayesian Neural Networks (BNNs) have shown promise in offering more interpretable model outputs with associated uncertainty estimates by leveraging variational inference and other approximations; however, this comes at a steep computational cost: at prediction, the network must make many forward passes to predict the posterior distribution of the outputs. This sampling of the network not only comes with significant computational cost, but also at an increasingly undesirable energy and environmental cost that drives end for improvements in this area. In statistics, the field of Computational Bayesian Inference has several methods on improving the speed of Bayesian Inference that haven’t yet been applied to BNNs.
In this PhD project, we aim to:
- Review the existing methods and frameworks of improving Bayesian Neural Network performance.
- Translate existing computational improvements in Bayesian Inference to Bayesian Neural Networks
- Explore novel optimisations of Bayesian Neural Network modelling
- Apply these methods to benchmark medical machine learning datasets to show improvements from prior methods