Bridging Domain Gap for Transfer Learning on Visual Tasks

Deep Neural Networks (DNNs) have achieved great performance in computer vision tasks. However, the performance of the DNNs would drop if the test dataset follows a distribution different from the training dataset. This issue is called domain shift. Another issue is that the DNNs need to be trained with a large amount of labeled data to avoid overfitting because of the large number of parameters. Collecting and labeling such a large volume of data is expensive and sometimes not possible. Transfer learning has emerged to tackle these challenges. It aims to extract knowledge from source problems to help learn other target problems in different domains and tasks. This thesis focuses on studying transfer learning from two different settings including domain generalization and unsupervised transfer learning.