Please send all enquiries via email (with CV) using the tag
[Phd-Enquiry] in the subject, along with either
[<project title>] or [
General] depending on whether you're applying to work on a specified project below or otherwise.
In reaching out to me, it would be quite helpful if you explain (briefly)
- why you want to work with me specifically,
- what you've worked on,
- what you'd be interested in exploring going forward, and
- how that would fit in with what I've been working on.
Outside of the potential opportunities listed below, there are a range of avenues for pursuing a PhD, including through the School of Informatics, or the Centres for Doctoral Training (CDTs) in Data Science, Natural Language Procesing (NLP), Robotics, and Biomedical AI.
Please refer to the relevant information on these web pages to learn about the application process, timescales, and funding options.
Application deadlines for starting a PhD in September 2021 are in late 2020 / early 2021. The sooner you get started, the better.
For prospective Masters students, please see the School's MSc by Research page for details on applying.
If you're already an Informatics MSc student looking for a thesis project, send me an email using the tag [
MSc-Project] in the subject.
I currently have the following two fully-funded positions for PhD students. Please note eligibility constraints for applicants.
Learning Structured Decompositions of Data for Interpretability
Funding: Edinburgh-Huawei Research Lab
Eligibility: UK/EU only
Start: Jan 2021
Perceptual data, such as vision or speech, is typically conceptualised in terms of structured compositions of atomic elements---for example objects and their spatial arrangements, or words and their sentential structures. Thus far, approaches have either assumed knowledge of explicit composition rules (e.g. grammars) and learnt the atomic elements of composition, or fixed the atomic elements and learnt to combine such elements implicitly (recurrent models), or done everything implicitly (transformers). This project will involve exploration of models and algorithms that can jointly learn explicit compositional structures for data, as well as learning the representations for the atomic elements employed within such compositions, such that things are rendered human-interpretable.
Interpretable Representation Learning through Embodied Constraints
Funding: School of Informatics
Start: Jan/Sept 2021
The goal of this project is to leverage data from multiple modalities, such as vision, language, and proprioception, to learn representations in deep latent-variable models that faithfully capture the conceptual commonalities across modalities, for example, what a kite is, or what it means to have wings, without explicit supervision, while retaining the abilities typically expressed within individual domains---interventions and generations. The idea is to reason about the characteristic latent features derived from such models---to identify interpretable components, capture notions of 'relatedness', along with broad dependency structure reflecting the nature of collective data observed.