Research

Learning as adynamical process.

We study how learning unfolds in machines and animals.

Sleeping bat representing the BATai research direction Awake bat representing the BATai research direction

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.

fairness data imbalance spurious correlations

How does bias evolve during training?

Which dataset and geometrical properties induce unfair solutions?

Can simple theory expose where mitigation strategies help or fail?

Awake llama representing the LLAMA research direction Sleeping llama representing the LLAMA research direction

Curricula, continual learning, transfer

LLAMA: Laws of Learning 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 in machine learning and cognitive science problems to understand 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 it does not?

How does curriculum affect the representations learnt by a neural network?

How do design choices like model parameterisation and initialisation affect forgetting?

Sleeping badger representing the BADGER research direction Awake badger representing the BADGER research direction

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 optimisation methods succeed.

optimisation landscape momentum

Which critical points can trap gradient-based algorithms?

How do optimisation dynamics interact with landscape geometry?

How do acceleration methods change dynamics and algorithmic thresholds?