Keynote Speakers

 

Maja Pantic

Maja Pantic is an AI Scientific Research Lead in Facebook and a Professor of Affective and Behavioral Computing at Imperial College London, working on machine analysis of human non-verbal behaviour and its applications to human-centric AI. Prof. Pantic published more than 350 technical papers in the areas of machine analysis of facial expressions, audiovisual analysis of emotions and social signals, and human-centered machine interfaces. She has more than 35,000 citations to her work, and has served as the Key Note Speaker, Chair and Co-Chair, and an organization/ program committee member at numerous conferences in her areas of expertise.


 

 

Lukas Wutschitz

Lukas Wutschitz is a senior researcher at the privacy preserving machine learning (PPML) group at Microsoft365 Research. His research interests include privacy preserving machine learning with a focus on large language models and natural language processing. He has worked on differentially private fine-tuning of language models, auditing of differential privacy guarantees, and quantifying privacy risk. Prior to joining Microsoft, he obtained his PhD in Physics from the Cavendish Laboratory at the University of Cambridge.


Title of  the Talk: Privacy preserving training of large language models


Abstract: 


Training large language models (LLMs) on sensitive data poses significant privacy challenges. In this talk, we will present a method to fine-tune LLMs of various sizes (up to 150b parameters) with mathematical guarantees of differential privacy (DP), a rigorous notion of privacy that ensures the protection of individual data subjects. We will explain the key concepts and properties of DP, such as composition and post-processing robustness, and how they can be applied to deep learning training. We will also demonstrate how to audit and verify the privacy guarantees of our method. Finally, we will introduce our research internship programme where much of the earlier presented work has originated.

 

 

Grigorios Pavliotis

Grigorios Pavliotis is Professor of Applied Mathematics at the Department of Mathematics at Imperial College. His main research interests lie in the areas of stochastic differential equations and diffusion processes, nonequilibrium statistical mechanics and homogenization theory for partial differential equations and stochastic differential equations. He is particularly interested in the development of analytical, computational and statistical techniques for multiscale stochastic systems, in time-dependent statistical mechanics and kinetic theory and in the analysis and development of sampling techniques in high dimensions. Current research projects include inference and control for multiscale systems, the development of computational techniques for calculating transport coefficients, homogenization for multiscale diffusion processes and sampling techniques in molecular dynamics.

Title of talk: Interacting particle systems and their mean field limit: propagation of chaos, phase transitions, fluctuations and inference

Abstract:

 

I will present recent results on the quantitative study of stochastic interacting particle systems and of their mean field limit. I will discuss about the links between uniform propagation of chaos, the absence of phase transitions and Gaussian fluctuations around the mean field limit. I will also present methodologies for inferring parameters in the mean field (McKean) stochastic differential equation from observations of particle paths.