Representation learning: from non linear decision to generative models
Statistical learning has evolved significantly in recent years, particularly around the concept of representation learning. After a quick introduction to present the key notions of robustness and generalization, we show how this paradigm has deeply transformed many applications by introducing semantics while allowing scaling. This first line of research is thus focused on data robustness and understanding with an application focus on sentiment analysis and recommender systems.
We then explore the contribution of end-to-end generative models to improve semantics but also to address new problems and better explain to the end user the origin of automated decisions. We approach this problem through a disentangling strategy to isolate and represent the factors allowing to reconstruct a signal but also to categorize it. To illustrate these approaches, we will once again use profiling applications and signals from the smart city.
Finally, we will present the perspectives offered by this work in the form of a research project for the years to come.
Phd defence : 02/17/2021
Jury members :
Massih-Reza Amini, Université Grenoble Alpes [Rapporteur]Patrice Bellot, Université Aix-Marseille [Rapporteur]
Gabriella Pasi, Université de Milan [Rapporteur]
Mohamed Chetouani, Sorbonne Université
Martin Trepanier, Université de Montréal
Emmanuel Viennet, Université Paris 13
Departure date : 02/17/2021