Annonce en lien avec l’Action/le Réseau : Doctorants

Laboratoire/Entreprise : CentraleSupélec SONDRA and ONERA
Durée : 3 years
Contact : jean-philippe.ovarlez@onera.fr
Date limite de publication : 2019-09-31

Contexte :
Radar signals are generally complex-valued (In-Phase and Quadrature channels with reduced Shannon sampling rate, polarimetric channels, interferometric channels, etc.). Also, radar processing schemes are generally based on complex filtering (FFT, Wavelets, Wiener, Matched Filter, etc.) and so impossible to be developed with classical Neural Network. Nowadays, Machine Learning Networks developed in the scientific community are mainly based on real nature of the signals (images, etc.). If the richness of information (mainly related to its physical meaning nature) contained in the phase has to be exploited, conventional Deep Neural Networks schemes have to be completely revisited.

Sujet :
We propose in this PhD topic to develop new architectures of Neural Network taking into account the the complex valued nature of radar signals. These new schemes will be based on the design of complex valued activation functions, complex thresholds, complex-valued optimization methodologies based mainly on complex gradient-descent-based problems. Finally, these new methodologies will be analyzed in terms of convergence of extended backpropagation algorithms (allowing the computation of complex neural weights). The improvement of such systems will be also analyzed in terms of performance compared to traditional Neural Networks.

Profil du candidat :
Master 2, High-level Engineer school

Formation et compétences requises :
Strong skills in Mathematics, Statistics, Statistical Signal Processing

Adresse d’emploi :
ONERA Palaiseau and CentraleSupelec,
Paris Saclay

Document attaché : Complex-Valued-DNN-for-Radar-Applications.pdf