Post-Doctoral Research Visit F/M Stochastic non-smooth optimization by first-order methods

Two-year postdoctoral position shared between the Inria GHOST team and the DAO team of the Laboratoire Jean Kuntzmann (LJK), funded by the Structured Optimization and Learning chair of the MIAI cluster and supervised by Mathieu Besançon (Inria, LIG), Quoc-Tung Le (UGA, LJK) and Jérôme Malick (CNRS, LJK). The project analyses adaptive optimization algorithms that tune hyperparameters such as the step size from data accumulated during optimization, including AdaGrad, RMSProp, Adam and related variants.

Most existing theory covers smooth functions; the aim is to extend it to stochastic, non-convex and non-smooth optimization, narrowing the gap between theory and the practical learning problems where these methods are already used. Research questions include defining a problem class close to learning models on which non-smooth algorithms can be analysed, studying AdaGrad-type algorithms in the non-smooth setting, and identifying necessary and sufficient conditions for adaptive algorithms to converge.

Collaborations with the Zuse Institute Berlin and the Institut Mathématique de Toulouse are possible. Start date 1 January 2027; gross salary €2,927 per month.