PROJECT TITLE :

Iterative Refinement for Multi-source Visual Domain Adaptation

ABSTRACT:

One of the most difficult aspects of multi-source domain adaptation is figuring out how to minimize the differences in domains that exist between each source domain and a target domain. After this step, you must evaluate the relevance of each source domain to the target domain in order to determine the appropriate amount of information to bring over from the various source domains. However, the vast majority of previously proposed methods barely take into account both discrepancies and relevance between domains. In this paper, we propose an algorithm that will solve the problem of semi-supervised domain adaptation using multiple sources. The algorithm's working title is Iterative Refinement based on Feature Selection and the Wasserstein distance (IRFSW). IRFSW's primary objective is to investigate the discrepancies as well as the relevance that exist between domains through the use of an iterative learning procedure. This procedure gradually improves learning performance up until the algorithm is terminated. During each iteration, we develop a sparse model for both the source domain and the target domain. This model is used to select features in which the domain discrepancy and training loss are simultaneously reduced. After that, a classifier is built using the features that were chosen from the source data and the target data that had been labeled. After that, we calculate the weights that have been transferred by employing optimal transport over the features that have been chosen. When combining the learned classifiers into an ensemble, the weight values are used as the ensemble weights. This allows the amount of knowledge that is transferred from source domains to target domains to be controlled. The effectiveness of the proposed method is demonstrated by the results of the experiments.