Chang Liu’s PhD Dissertation Defense

“Unleashing the Potential of Transfer Learning for Visual Applications”

Committee Members:
Prof. Raymond Fu (Advisor)
Prof. Sarah Ostadabbas
Prof. Zhiqiang Tao

Abstract:
The recent flourish of deep learning in various tasks is largely accredited to the rich and accessible labeled data. Nonetheless, massive supervision remains a luxury for many real-world applications. Further, the domain shift problem has also seriously impeded large-scale deployments of deep-learning models. As a remedy, Transfer learning aims at improving the performance of target learners on target domains by transferring the knowledge contained in different but related source domains. In this way, the dependence on a large number of target domain data can be reduced for constructing target learners.

In this dissertation research, I investigate two major problems in transfer learning, domain adaptation (DG) and domain adaptation (DA), on various visual applications. (1) The challenge of DG lies in an over-simplified assumption, that is, the source and target data are independent and identically distributed (i.i.d.) while ignoring out-of-distribution (OOD) scenarios commonly encountered in practice. This issue is common in visual applications such as object recognition, hyperparameter optimization, and face recognition. We propose algorithms that are specifically designed for each task, such as metric learning, adversarial regularization, feature disentanglement, and meta-learning. (2) DA can be considered a special case of DG with unlabeled target data available. The major challenge is how to align the labeled source and unlabeled target data. We delve into the applications of image recognition and video recognition and propose algorithms to ensure domain-wise discriminativeness and class-wise closeness across domains. Experiments show that the proposed algorithms outperform the state-of-the-art methods on the commonly-used benchmark datasets.