PROJECT TITLE :

Unsupervised Domain Adaptation via Discriminative Manifold Propagation

ABSTRACT:

It is possible to successfully leverage rich information from a labeled source domain into an unlabeled target domain through the use of unsupervised domain adaptation. Despite the fact that deep learning and adversarial strategies have made a significant breakthrough in the adaptability of features, there are still two issues that need to be studied in greater depth. First, the target domain's hard-assigned pseudo labels are arbitrary and prone to error; furthermore, the direct application of these labels has the potential to destroy the intrinsic data structure. Second, the training of deep learning algorithms using batches restricts the ability to characterize the overall structure. To accomplish both transferability and discriminability at the same time, the authors of this paper propose a learning framework based on Riemannian manifolds. This framework establishes a probabilistic discriminant criterion on the target domain by making use of soft labels. This addresses the first problem. This criterion is extended to a global approximation scheme for the second issue. It is based on prototypes that have already been pre-built. In order to maintain compatibility with the embedding space, the manifold metric alignment was chosen. For the purpose of providing constructive guidance, the theoretical error bounds of various alignment metrics have been derived. The method that has been proposed can be utilized to solve a number of different varieties of domain adaptation problems, which include settings that are both vanilla and partial. Extensive experiments have been carried out in order to look into the method, and a comparative study demonstrates that the discriminative manifold learning framework is superior.