Research themes

AI moves faster than any one group can keep up with. This theme is how we try: reading the new work, testing it, and forming a view on which parts of it matter for us.

Most of our interest sits in agentic systems that plan and act, in learning from feedback, and in large models used as scientific instruments rather than products.

Trust runs through all of it. If a system is going to be relied on, someone has to be able to show why. That question turns up in every other theme on this page.

Health is a hard place to do machine learning well. Records are fragmented and unevenly kept. The people the data came from are often not the people a model ends up being used on. The bar for evidence is higher than in most fields, and it should be.

So accuracy is only part of the problem. How does a model behave on patients it never saw in training? What would a clinician need to see before acting on what it says? Those questions interest us as much as the modelling does.

Archives are partial. Language shifts under you. Sources contradict each other. Treat all of that as noise to be cleaned away and you usually delete the thing that made the question worth asking.

The traffic runs both ways. Computation can open up collections far larger than anyone could read. Humanities scholarship, in return, is very good at pressing on provenance, interpretation and bias, which machine learning tends to leave unexamined.

Natural science already has good theory, and the point is not to replace it. Learning earns its place where the theory is expensive to run, incomplete, or hard to reconcile with the measurements you can actually take.

That makes the constraints part of the research problem. A surrogate model can be fast, fit the data, and still break conservation of energy. It is then useless as a scientific instrument, whatever its error numbers say.

Engineering already runs on models: simulations, digital twins, optimisation problems, control laws. Every one of them depends on a person who knows the domain well enough to choose the right simplifications.

We are interested in what changes when an AI system takes part in that loop instead of watching from outside it. Choosing what to model. Judging which assumptions hold. Working out why something behaves the way it does.

What matters in the end is the decision, not the prediction. A model can be accurate and still push a design in the wrong direction.