Description
Date depot: 20 mai 2022
Titre: Deep Learning for Modeling Dynamical Systems
Directeur de thèse:
Abdenour HADID (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Deep learning offers a new data driven approaches to the modeling of dynamical systems underlying natural observations. This has recently given rise to new and prolific research topics focused on exploiting deep learning methods for modeling spatio-temporal dynamics. The use of deep learning and data-driven approaches for modeling natural phenomena from physical observations however suffers from limitations such as the lack of generalization, robustness and physical plausibility. Modeling natural phenomena from physical observations indeed raises new challenges for machine learning and deep learning. Being the result of multiple interacting physical processes, the observed phenomena can be extremely complex. The data are heterogeneous, noisy, and even when plentiful they usually represent scarce and partial information about the underlying process. So, the main question which arises is: how to build accurate and fast machine learning models from this limited training data. Recent studies showed that Transforms can outperform conventional deep learning models in many applications such as natural language processing, computer vision, and audio processing. A transformer is basically a deep learning model that adopts the mechanism of attention, weighing the influence of different parts of the input data. It is hence appealing to adapt the mechanism of Transformers and beyond it to our specific downstream tasks and applications.
Doctorant.e: Sasal Léna