High-Dimensional Asymptotics and Dataset Selection for Private Transfer Learning
External datasets can improve prediction, but distribution shifts and privacy noise can erase those gains. This work develops...
We use statistical physics, probability, and computer science to understand why learning systems generalise, fail, and develop bias, and how to build safer and more intelligent AI.
Explore our researchOur latest works
External datasets can improve prediction, but distribution shifts and privacy noise can erase those gains. This work develops...
Drug discovery teams commonly summarise a dose–response experiment by the concentration at which a compound produces half its...
News
Our work in the news! Business Day features our ICML 2026 position paper on model collapse and the risks it poses to low-resource communities. The article by Andile Masuku and Damilare Dosunmu includes comments from Devon Jarvis and focuses on the implications for African languages.
Stefano will present Curriculum learning in humans and neural networks as a plenary speaker at the 3rd Nordic Biomathematics Meeting in Uppsala, Sweden, on 3–4 December 2026.
Stefano is in London for the PIAMI workshop on 16–18 November, joining the panel on Limits of theory and the methodological frontier.
How much can a single potency estimate tell us about a drug candidate? In a ChemRxiv preprint from his master’s project with AstraZeneca, Anton O’Nils worked with Stefano, Ola Engkvist, Jens Sjölund and Annie M. Westerlund to predict the full dose–response curve from molecular structure. They tested it on compounds screened for inhibition of hERG, a cardiac ion channel. Predicting the full curve gave more accurate potency estimates than the scalar models they compared.
We work across machine learning, statistical physics, cognitive science, and high-dimensional probability at Chalmers and the University of Gothenburg, with active links to Wits.