Research

Learning as a dynamical process

We study how learning unfolds in machines and animals: when training amplifies structure in the data, when curricula and transfer help, and how optimisation dynamics move through high-dimensional landscapes.

BATai Bias Amplification and Transfer in AI

Fairness, imbalance, and transient learning dynamics.

LLAMA LLearning in Animals and MAchines

Curricula, shaping, transfer, and continual learning.

BADGER Bridging Algorithm Dynamics with GEometRy

Optimisation, landscapes, thresholds, and trajectories.

Bias, imbalance, fairness

BATai: Bias Amplification and Transfer in AI

Machine-learning systems are often trained on data that already contain imbalance, hidden structure, and historical bias. We build solvable models that isolate how these ingredients shape learning trajectories and fairness outcomes.

The aim is to make bias formation mechanistic: which design choices amplify disparities, which mitigation strategies trade off accuracy and fairness, and when supposedly neutral training choices are not neutral at all.

How does bias evolve during training, not just at convergence?

Which geometries of data imbalance induce unfair solutions?

Can simple theory expose where mitigation strategies help or fail?

Curriculum learning and overparameterisation

Curricula, continual learning, transfer

LLAMA: LLearning in Animals and MAchines

Animals learn through shaping, curricula, transfer, and continual exposure to related tasks. Modern neural networks can behave very differently, even when trained on problems that look simple to biological learners.

We use controlled models and collaborations across machine learning and cognitive science to ask when training order matters, why networks forget, and how artificial learners can become better computational models of animal learning.

curriculum learning continual learning transfer learning reinforcement learning

When does a curriculum speed learning, and when does it disappear?

Why does intermediate task similarity often create the most forgetting?

Can shaping reveal reusable primitives for compositional tasks?

Optimisation, landscapes, geometry

BADGER: Bridging Algorithm Dynamics with GEometRy

In high dimensions, intuition about optimisation often fails. A random starting point, a training algorithm, and the geometry of the loss landscape together determine which solution learning can actually find.

We connect dynamical descriptions of algorithms with geometric descriptions of landscapes to understand thresholds, spurious minima, slowdowns, and the regimes where descent methods succeed.

optimisation landscape momentum

Which critical points can trap descent algorithms?

How do algorithmic trajectories interact with landscape geometry?

When do acceleration methods change dynamics without changing thresholds?

Common language across projects

The projects differ in topic, but share a way of working: reduce the learning problem until it is analyzable, then test which lessons survive in richer models and data.

Solvable models

Teacher-student systems, Gaussian mixtures, random features, and high-dimensional limits.

Learning dynamics

Exact trajectories, transient behaviour, forgetting curves, and protocol design.

Geometry

Loss landscapes, local minima, thresholds, and data manifolds that shape algorithms.

Empirical grounding

Numerical experiments, neural-network simulations, cognitive tasks, and real datasets.