to more general settings and understanding why it works
#Optimisation et théorie des jeux pour l’IA
The main objectives are to extend the scope of algorithms that can cope with nonconvexity and the curse of dimensionality by exploiting data information and structures, to analyze their mathematical properties, to identify the determining factors of their numerical complexity, to improve their performance, and to apply them to solve high impact applied problems.
Many ML problems require solving nonconvex optimization, e.g., clustering problems; dimension reduction paradigms such as sparse PCA (Principal Component Analysis), Nonnegative Matrix Factorization, to name just a few. In all these instances, the problems are highly nonconvex, huge scale and even nonsmooth, and are the source of challenging open questions.