As machine learning approaches have been applied to more challenging problems, there has been increasing interest in understanding both how decisions have been made, interpretability, and levels of confidence, uncertainty, associated with decisions. Both these tasks are highly challenging for deep-learning based approaches, which rely on highly distributed non-linear representations.
This project focuses on the second problem, how to enable systems to automatically generate a measure of confidence in a decision and the cause of this uncertainty. One solution to this problem is to adopt a Bayesian approach where samples are drawn from the posterior distribution over the model parameters, and attributes of the set of predictions from this ensemble used to infer information about prediction uncertainty.
Though theoretically sound, Bayesian approaches are challenging to apply deep learning models which have millions, or tens of millions, of parameters. Recently an approach called prior networks has been proposed that aims to directly model the distribution over the predictions of the ensemble. The complexity of this technique scales with the number of classes, rather than model parameters. To date these approaches have been applied to tasks such as image classification.
This project aims to investigate how they can be applied to structured data tasks, using applications is speech and language processing to evaluate the approaches. The additional challenge in these application areas is that errors can occur in bursts. For example if a speaker is talking in an environment with high background noise then all the words in that utterance will have high levels of uncertainty.
Applicants should have (or expect to obtain by the start date) at least a good 2.1 degree in an Engineering or related subject.
EPSRC DTP studentships are fully-funded (fees and maintenance) for UK students or provide fees only for EU students from outside the UK. Further details about eligibility can be found at: https://epsrc.ukri.org/skills/students/help/eligibility/.
Applications should be submitted via the University of Cambridge Applicant Portal www.graduate.study.cam.ac.uk/courses/directory/egegpdpeg, with Prof Mark Gales identified as the potential supervisor.
The University values diversity and is committed to equality of opportunity.