Random matrix theory and the high-dimensional optimization dynamics of machine learning
Gérard Ben Arous
Courant Institute of Mathematical Sciences, NYU
https://cims.nyu.edu/~benarous/
Date(s) : 16/05/2025 iCal
16h00 - 17h00
I will survey recent progress in the understanding of the optimization dynamics for Machine Learning classical classification tasks (for the central case of Gaussian mixture models) via the spectral theory of their Hessian and information or Fisher matrices along the algorithm path. In particular we will see how a classical spectral transition of Random Matrix Theory plays a major role in this context.
This relies on joint recent work with Reza Gheissari (Northwestern), Jiaoyang Huang (Wharton) and Aukosh Jagannath (Waterloo).
Emplacement
Saint-Charles - FRUMAM (2ème étage)
Catégories
