The CSP-IAS Institute for Advanced Study of AI4I opens its 2026 seminar programme with an event dedicated to one of the most pressing structural questions in contemporary artificial intelligence: how to train neural networks effectively as scale increases.
IAS Seminar #01/2026 will take place on 19 January 2026, from 10:00 to 12:00 CET, at OGR Torino – Sala Mezzanino, and online. The seminar is part of the IAS programme developed by AI4I in collaboration with the Fondazione Compagnia di San Paolo, with the aim of fostering advanced scientific discussion at the intersection of theory, computation and industrial impact.
The seminar, titled “Training Neural Networks at Any Scale”, will be led by Volkan Cevher, Associate Professor at the École Polytechnique Fédérale de Lausanne and Faculty Fellow in the Electrical and Computer Engineering Department at Rice University.
At the core of modern deep learning lies the notion of scale: ever-growing datasets, increasingly powerful computational infrastructurs and neural architectures designed to exploit them. Yet scaling brings with it critical challenges, including training instability, rising costs and the need for extensive model-specific tuning. As large-scale training requires substantial resources, the role of theory becomes essential in formulating reliable hypotheses and guiding design choices.
Cevher’s seminar explores the mathematical foundations underlying scaling in deep learning, with a particular focus on the numerical optimization algorithms commonly employed across application domains, from computer vision to language models. By presenting these methods within a unified conceptual framework, the talk highlights the interplay between online learning, adaptation to smoothness structures and optimization geometry through non-Euclidean norms. Rather than simply advocating larger models, the seminar offers a perspective on strategic scaling, aimed at improving robustness and efficiency while economising on computational resources.
The IAS Seminars are conceived as a continuous scientific space within AI4I: a programme designed to host leading international researchers, encourage rigorous debate and contribute to the long-term development of artificial intelligence beyond short-term trends. The 2026 programme continues this trajectory, reinforcing the role of the IAS as a platform for advanced research and shared reflection.
Programme – 19 January 2026
10:00 – Attendee registration
10:30 – Seminar
11:30 – Q&A
12:00 – End of seminar
The event will be held in person and online.
