We are pleased to announce that Lenka Zdeborová will give a lecture as part of the Munich AI Lectures Series, presenting recent theoretical advances in the analysis of attention mechanisms, the core building blocks of modern machine learning architectures.

This lecture focuses on the theoretical understanding of attention layers in high-dimensional regimes, addressing how such models learn and generalize from data. Using analytically tractable models, learning dynamics and generalization performance are characterized in closed form for the first time, offering precise predictions and mechanistic insight into the representational power of attention-based architectures.

The talk will concentrate on supervised learning settings, where these models allow for rigorous theoretical analysis. In addition, the lecture will outline how these results open the door to manageable theoretical frameworks for self-supervised and generative training with attention mechanisms.

Topics will include:

• Theoretical analysis of attention layers in high-dimensional models

• Generalization behavior in token-based learning tasks

• Mechanistic insights into representation learning with attention

• Implications for self-supervised and generative models

About the speaker
Lenka Zdeborová is Professor of Physics and Computer Science at EPFL, where she leads the Statistical Physics of Computation Laboratory. Her research bridges statistical physics, machine learning, inference, and optimization, with a focus on explaining how modern AI systems learn, generalize, and scale. She is the recipient of several prestigious awards, including the CNRS Bronze Medal, the Irène Joliot-Curie Prize, and multiple ERC Grants.

Date & Time
Thursday, January 29, 2026
03:00-04:00 PM (Europe/Berlin)

Location
Munich AI Lecture
Theresienstr. 90, Room 0101.Z1.090
80333 Munich, Germany

More information here