Documents disponibles écrits par cet auteur

Graph neura network / Mohamed Mehdi Atmana 

Public ISBD Titre : Graph neura network : a comparative study Type de document : document électronique Auteurs : Mohamed Mehdi Atmana, Auteur ; Ibrahim Laiche, Auteur ; Rachid Illoul, Directeur de thèse ; Asma Atmana, Directeur de thèse Editeur : [S.l.] : [s.n.] Année de publication : 2020 Importance : 1 fichier PDF (3.2 Mo) Présentation : ill. Note générale : Mode d'accès : accès au texte intégral par intranet. Mémoire de Projet de Fin d’Études : Automatique : Alger, École Nationale Polytechnique : 2020 Bibliogr. F. 93 - 100 Langues : Anglais (eng) Mots-clés : Machine  learning Graph  neural  network Attention  mechanism Supervised  learning Graph  classification Index. décimale : PA00120 Résumé : Graphs are a powerful data structure that is used to model objects and their interactions. Owing to their ability to capture rich information about interacting entities, they are used to model a wide range of data. For example, in robotics, an articulated robot's bodies can be represented with a graph's nodes while the joints linking the bodies together can be represented with a graph's edges. Recently, neural architectures that are able to process graph input data, called Graph Neural Networks(GNNs), haveemerged with promising results on many supervised learning tasks such as graph classification. Work has also been done to use these GNNs in many control-related applications such as for inference, system identification, model-predictive control and even for visual tasks like human action recognition. In this work, after an extensive review of the state-of-the-art literature on GNNs, including breakthrough applications in control, we combine two reference GNN architectures with a powerful attention mechanism, proposing two novel architectures and validating them on four benchmark graph classification datasets using rigorous methodology and cutting-edge software tools. Our proposed architectures achieve impressive gains in performance of up to 14% over baselines on certain datasets, opening up interesting perspectives for future work, especially in human action recognition and pose estimation problems. Graph neura network : a comparative study [document électronique] / Mohamed Mehdi Atmana, Auteur ; Ibrahim Laiche, Auteur ; Rachid Illoul, Directeur de thèse ; Asma Atmana, Directeur de thèse . - [S.l.] : [s.n.], 2020 . - 1 fichier PDF (3.2 Mo) : ill. Mode d'accès : accès au texte intégral par intranet. Mémoire de Projet de Fin d’Études : Automatique : Alger, École Nationale Polytechnique : 2020 Bibliogr. F. 93 - 100 Langues : Anglais (eng) Mots-clés : Machine  learning Graph  neural  network Attention  mechanism Supervised  learning Graph  classification Index. décimale : PA00120 Résumé : Graphs are a powerful data structure that is used to model objects and their interactions. Owing to their ability to capture rich information about interacting entities, they are used to model a wide range of data. For example, in robotics, an articulated robot's bodies can be represented with a graph's nodes while the joints linking the bodies together can be represented with a graph's edges. Recently, neural architectures that are able to process graph input data, called Graph Neural Networks(GNNs), haveemerged with promising results on many supervised learning tasks such as graph classification. Work has also been done to use these GNNs in many control-related applications such as for inference, system identification, model-predictive control and even for visual tasks like human action recognition. In this work, after an extensive review of the state-of-the-art literature on GNNs, including breakthrough applications in control, we combine two reference GNN architectures with a powerful attention mechanism, proposing two novel architectures and validating them on four benchmark graph classification datasets using rigorous methodology and cutting-edge software tools. Our proposed architectures achieve impressive gains in performance of up to 14% over baselines on certain datasets, opening up interesting perspectives for future work, especially in human action recognition and pose estimation problems.

Exemplaires

Code-barres	Cote	Support	Localisation	Section	Disponibilité	Spécialité	Etat_Exemplaire
EP00055	PA00120	Ressources électroniques	Bibliothèque centrale	Projet Fin d'Etudes	Disponible	Automatique	Téléchargeable

Documents numériques

ATAMNA.Mohamed-Mehdi_LAICHE.Ibrahim URL