Today starts the Mathematical Foundations of AI workshop, organised by Flavio and Stefano. The programme brings together diffusion models, transformers, and associative memories, and includes a contributed talk by Chenxiao on our work on heterogeneous data in diffusion models.
Chalmers - University of Gothenburg - Wits
We build theoretical foundations for learning in machines and animals.
Our work focuses on bias, generalisation, curriculum learning, continual learning, and optimisation dynamics.
Our research questions
Our projects connect exact models, experiments, and modern machine learning systems to understand when learning succeeds, fails, or amplifies structure in data.
Recent activities
Talks, papers, workshops, and funding updates from the group.
Great news! Stefano has been awarded 39,000 SEK from Helge Ax:son Johnsons stiftelse, supporting research travel and international scientific exchange.
Stefano is giving a talk on curriculum learning at BrainNet 2026 on May 27.
New preprint from the group! Flavio, Chenxiao, Enrico, Luca, and Stefano released The Interplay of Data Structure and Imbalance in the Learning Dynamics of Diffusion Models, studying how class structure and imbalance shape generalisation and memorisation in diffusion models.
Recent publications
A short view of the newest papers. The full list includes work on statistical physics, curricula, fairness, diffusion models, and optimisation.
arXiv preprint
The Interplay of Data Structure and Imbalance in the Learning Dynamics of Diffusion Models
Real-world datasets differ across classes in both structure and frequency, but most theory for diffusion models assumes homogeneous data. This work develops a high-dimensional analytical framework for class-dependent...
ICML 2026 (Spotlight)Position: the Stochastic Parrot in the Coal Mine. Model Collapse is a Threat to Low-Resource Communities
Model collapse can degrade generative models when they are trained on outputs from earlier models. This position paper argues that the problem compounds existing concerns around large language...
ICML 2026Sharp description of local minima in the loss landscape of high-dimensional two-layer ReLU neural networks
We study the population loss landscape of two-layer ReLU networks in a realisable teacher-student setting with Gaussian covariates. The work shows that local minima admit an exact low-dimensional...
The group
A small team working across theory, machine learning, cognitive science, and statistical physics.