Can Qin’s PhD Proposal Review

“Transfer Learning across Domains, Tasks and Models”

Abstract:

The big data stands as a cornerstone of deep learning, which has significantly improved a wide range of machine learning and computer vision tasks. Despite such a great success, data collection is time-consuming and costly, considering manual efforts and privacy restrictions. Transfer learning is a promising direction toward data-efficient AI by leveraging acquired data and pre-trained models as guidance. This dissertation focus on the feature and model transfer across different domains and tasks, which can be roughly summarized into three sections. (1) Section One focuses on Unsupervised Domain Adaptation (UDA) without any labels in the target domain. The technical challenge of UDA is the distribution mismatch across domains. I have presented a hierarchical alignment model as the solution. (2) Section Two extends UDA into semi-supervised domain adaptation (SSDA) with minimal target-domain labels, which is useful and effortless to acquire. To avoid overfitting toward labeled data, I have proposed structural regularization verified on different classification benchmarks. (3) Section Three mainly explores the model transfer, including teacher-student knowledge distillation and heterogeneous models infusion with a high potential for model compression and enhancement.

Committee:

Prof. Raymond Fu (Advisor)

Prof. Octavia Camps

Prof. Huaizu Jiang