Research

Four pillars, and room for yours.

The lab works across four areas of artificial intelligence — and takes students from any academic background who want to bring their own field into one of them.

01 — Areas

Where the work sits.

These four areas are not separate departments. Most projects in the lab touch two or three of them at once — a generative model is only interesting here if someone is also asking who it works for and what it is good for.

Area 01
Modalities Shared

Multimodal AI

People do not experience the world one channel at a time. We read a face and a tone of voice together; we understand a photograph partly through the caption sitting under it. Multimodal AI is the attempt to build systems that do the same — models that hold vision, text and audio in one representation rather than treating each as a separate problem.

This is the lab's centre of gravity. Most of our projects are, underneath, an exercise in getting two modalities into a shared space and then generating one from the other: images into music, text into sequential images, visual stimuli into audio.

In practice

  • Visual question answering. Answering questions about an image, which requires grounding language in what is actually visible.
  • Image captioning. Producing description that is specific to the image rather than plausible for any image.
  • Cross-modal alignment. Learning embedding spaces where related content in different modalities lands in the same neighbourhood.
  • Cross-modal generation. Conditioning a decoder in one modality on an encoder in another — the shape of three of the four current projects.
Vision–Language Models Vision Transformers CNNs Contrastive learning Embedding alignment
Image-to-Music → Text-to-Comic → Gene Editing → VLM survey →
Area 02
Group A Group B Disparity

AI Ethics & Bias

A model that performs well on average can still fail badly and consistently for a particular group of people. Aggregate accuracy hides this by design. This area is about the measurement work that surfaces those failures, and the design choices that stop a system shipping with them intact.

We treat fairness, accountability and transparency as engineering requirements rather than a closing paragraph. That means disaggregated evaluation, documented datasets, and being specific about who a model is likely to disadvantage before anyone claims a result.

In practice

  • Disaggregated evaluation. Reporting performance per group, not only in aggregate, and treating the gap as a headline number.
  • Representation analysis. Auditing what is present and absent in training data, and in the media that data is drawn from.
  • Transparency. Documenting data provenance, model limitations and known failure modes alongside results.
  • Generative harms. What generative systems reproduce about people when nobody has asked them to.
Fairness metrics Dataset auditing Model cards Cross-cultural analysis Accountability
International collaboration → Media bias study →
Area 03
Noise Sample

Generative Models

Large language models and diffusion systems changed what a small lab can attempt. A team of two students can now build something that produces text, images, audio or code — which moves the interesting research away from "can it generate" and towards control, faithfulness and consistency.

That is where our questions sit. Not whether a model can produce a plausible panel, but whether the character in it is the same person as in the previous panel. Not whether it can produce pleasant music, but whether the music has anything to do with the image it was given.

In practice

  • Conditional generation. Steering output with an input from another modality, and measuring whether the conditioning actually took.
  • Consistency across sequences. Holding identity, style and continuity steady over many generated outputs.
  • Architecture study. Comparing CNNs, RNNs, transformers and diffusion approaches on the same task rather than assuming the newest wins.
  • Evaluation design. Building protocols that separate output quality from appropriateness to the input.
LLMs Diffusion models Transformers RNNs Conditioning Human evaluation
Text-to-Comic → Storytelling & T2A →
Area 04
Applied domains

AI for Social Good

The same techniques that generate a comic panel can extract drug interactions from clinical literature. This area is about pointing the lab's methods at problems where a better model produces a materially better outcome for someone — in healthcare, education and environmental work.

It is also where the lab's cross-disciplinary intake pays off most directly. A model built to analyse medical text is only as good as the domain knowledge shaping what counts as a correct answer, and that knowledge usually comes from a member who studied something other than computer science.

In practice

  • Healthcare. Information extraction from biomedical literature — the lab's drug–drug interaction work sits here.
  • Education. Language learning analytics, and models that adapt to how an individual learner is actually progressing.
  • Environment. Applying vision and sequence models to sustainability and monitoring problems.
  • Security. Understanding what generative AI changes about the threat landscape, and what it changes about defence.
Information extraction Biomedical NLP Learning analytics Applied ML
Drug–drug interaction paper → Multimodal gene editing → Language learning analytics →

02 — Methods

What runs underneath all four.

Vision

CNNs & ViT

Convolutional networks and Vision Transformers for encoding images and video into representations other models can use.

Language

Transformers & LLMs

Sequence models for understanding, extraction and generation — from token classification to open-ended text.

Generation

Diffusion & RNNs

Iterative denoising for images and audio, recurrent architectures where sequence order carries the meaning.

Practice

Python, PyTorch, TensorFlow

The working stack. Members implement from papers, run experiments and keep results reproducible.

03 — Any background

You do not need a computer science degree.

Students from every academic background are welcome here, and the lab is deliberately built that way. Artificial intelligence is a set of methods, not a subject — the questions worth asking usually come from somewhere else. A linguist knows what a meaningful segment boundary is. A medical student knows which extracted interaction is clinically significant. A designer knows when a generated page has stopped reading as a page.

What you bring is the domain. We teach the AI. Members arriving without a programming background start with the implementation fundamentals under a mentor and join an active project from the beginning, rather than waiting until they feel qualified — which, in research, never quite happens.

Several of the lab's published papers came from members whose degree was in something other than computer science. If your field has data, text, images, sound or human behaviour in it, there is a project here.

Your field × AI

Languages & Linguistics
Segmentation, morphology and learner-error analysis — deciding what a model should treat as a unit of language, and what counts as progress in learning one.
Project [08]
Medicine & Life Sciences
Extracting structured knowledge from biomedical literature, and judging which extracted relations are actually clinically meaningful.
SSRN · bioRxiv
Business & Management
AI-driven analytics for strategy, growth and operations — framing the decision before choosing the model.
Project [03]
Media & Communication
Representation and narrative at scale: what patterns appear across a corpus of media, and what a generative system reproduces from it.
Projects [04] [06]
Art, Music & Design
Cross-modal generation — image to music, script to comic page. Compositional judgement is the part a model cannot supply for itself.
Projects [09] [10]
Education
Learning analytics and personalisation — modelling how a student progresses, and what an intervention should look like.
Project [08]
Security & Systems
Vulnerability analysis in virtualised infrastructure, and what large language models change about both attack and defence.
Projects [01] [02]
Psychology & Social Science
Designing the human evaluation studies that decide whether a generated output is any good, and interpreting what the results mean.
All areas
What we ask

Willingness, not credentials.

You should be prepared to learn Python and to stay with an experiment past the first failed run. Prior machine learning experience helps but is not a filter — the mentorship exists precisely to close that gap.

What you get

A question of your own.

If you arrive with a problem from your field, we will work with you to shape it into something answerable and put a team around it. That is how several of the lab's projects started.

Bring your field

Tell us what you'd want to study.

Write to us with your background and the question you keep coming back to. If it can be approached with the methods above, we will tell you honestly — and if it can't, we will tell you that too.

How to apply → Current projects