Description
Date depot: 18 novembre 2019
Titre: Generative modelling and discriminative learning with invertible neural networks
Directeur de thèse:
Patrick GALLINARI (ISIR (EDITE))
Directeur de thèse:
Alain RAKOTOMAMONJY (LITIS)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé:
Invertible neural networks (INNs), which define a bijective mapping between distributions, have recently become a competitive approach w.r.t. to non-invertible counterparts both for discriminative and for generative tasks. Two complementary concepts have recently emerged for the design and analysis of INNs. One is the family of analytical invertible models known as Normalizing Flows, and another one relies on numerical inversion schemes developed for ordinary differential equations numerical schemes (ODE)
The thesis will explore these two approaches and develop the links between normalizing flows and ODEs for NN analysis and design. The initial objective is to focus on the following main directions:
- Theory and Algorithms:
-- Latent-variable inference and log-likelihood evaluation
-- Stability, Robustness and Invertibility guarantees
-- Design of new expressive NN formulations for classification or generation
- Applications. On the application side, the thesis will focus on:
-- Unpaired cross-domain translation. Applications of these ideas will be examined for domain adaptation.
--Time sequence modelling. Potential relevant applications concern recommendation tasks such as next product recommendation based on sequences of products.
Doctorant.e: Kirchmeyer Matthieu