Description
Date depot: 8 avril 2020
Titre: Extensions and Applications of Graph-based Neural Networks
Directrice de thèse:
Maria TROCAN (LISITE)
Encadrante :
Patricia CONDE-CESPEDES (LISITE)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
votre projetNetwork-based analysis is a relatively recent and broad field of study. Indeed, researchers from many disciplines are now more than ever involved with the collection, modeling and analysis of network-indexed data. Undeniably, many phenomena arising from different situations can be represented using complex networks. For instance, in computer science the internet network; in biology, the protein-protein interaction (PPI) networks; in sociology, the social networks; and in medicine the disease propagation in human-interaction networks for infectious diseases or the genetic relationship networks for rare diseases. Graphs are mathematical representations of networks.
More recently, artificial intelligence approaches have become a very active and promising research field since there is now increasingly more data and scalable storage, memory and computing capacity to run efficient and effective algorithms on available data. More specifically, the branch of machine learning called deep learning revolutionized the tools commonly used in mainly two fields, computer vision and natural language processing (NLP). Deep learning techniques distinguish from typical statistics tools on the type of input data they treat, non-structured data, such as image, text or signals.
Unlike images and text, there is an increasing number of applications where data are represented as graphs with complex relationships and interdependency between objects. The complexity of graph data has imposed significant challenges on existing deep learning algorithms since graph data require robust approaches tailored to capture their structure. Graph Neural Network (GNN) is the branch of machine learning which concerns on building neural networks for graph data. Despite the potential applications of GNN, only recent studies exist about applying deep learning methods to graphs [1,2]. Many practical situations justify the need of using GNN. In computer vision, recognizing semantic relationships between objects facilitates the understanding of the meaning behind a visual scene; identifying human actions using Human joints which naturally form a graph. In addition, images often contain multiple objects, understanding the relationships (i.e., visual relationships) among them might be useful. In Natural language processing, text classification needs to take into account the inter-relations of documents or words to infer document labels and create a knowledge graph.
Doctorant.e: Lachaud Guillaume