Date/heure
3 mars 2021
14:00 - 15:00
Oratrice ou orateur
Pablo Jimenez Moreno (CMAP - Centre de Mathématiques Appliquées - Ecole Polytechnique)
Catégorie d'évènement Séminaire des doctorants
Résumé
Stochastic Approximation is a useful tool for Machine Learning techniques such as Stochastic Gradient Descent. These algorithms are applied to a lot of different fields, improving the transportation times, helping doctors diagnosing with medical images, automatically translating text, detecting spam and more. Most of the time, the model traditionally lies in a vector space. However, some problems present non-linear constraints, that can be translated into a manifold. This framework ensures the conservation of key properties. As an introduction to geometric machine learning, we study the gradient descent algorithm, and its adaptation to Riemannian manifolds. Finally, we compare the performance of the two, introducing new non-asymptotic bounds.