Etudiant : GHANNOU OMAR

Filière : Master Web Intelligence et Sciences des Données (WISD)

Encadrant : Pr. YAHYAOUYI ALI

Annèe : 2022

Résumé : Multi-source Domain Adaptation (MDA) keeps the part of adapting models trained on multiple labelled source domains to unlabelled target domain. In this report we will introduce our approach as a collaborative MDA framework. It consists of two adaptation phases. Firstly, we make the domain adaptation for each source with the target individually, using the optimal transport, then as a second phase we define the architecture of the centralised federated learning to collaborate the N models that represent the N sources where this architecture enables the advantage of using the sources without access to the data, which resolve the sources data privacy issues in the domain adaptation. This final part of the framework and the second phase of the adaptation is where the server guide the adaptation and fine-tune it with a small number of pseudo-labelled samples available in the target domain, that we name it, the target validation part of the dataset.