Speaker: Mr Zijian Wang
Host: A/Prof Helen Huang

Seminar Type:  PhD Confirmation Seminar


With the increasing of data volume and computational power, deep learning has started to demonstrate its tremendous potential to the world over the last decade. Unfortunately, the scarcity of the annotated data become a significant barrier for the deep learning applications. Domain adaptation (DA) has therefore arisen to address such limitations by performing the knowledge transfer from a label-rich domain to a label-scarce domain. However, most of the existing DA methods assume the source and target data sharing exactly the same classes of objects or only enabling the knowledge transfer in a single modality. These assumptions strongly limit the usability of DA in real applications.

In this research, we are focusing on developing effective knowledge transferring algorithms with a relaxed assumption to enhance the applicability. At the first stage, we investigate a practical problem of open-set domain adaptation and designed an end-to-end Progressive Graph Learning (PGL) framework. To consider the multi-modality knowledge transfer, we propose the Prototype-Matching Graph Network (PMGN), which facilitates the knowledge transfer between heterogeneous domains by gradually exploring the domain-invariant class prototype representations. Extensive experiments demonstrate the superiority of our methods over the state-of-the-arts in open-set domain adaptation and heterogeneous domain adaptation.


Zijian Wang received the B.E. (Hons.) degree from the University of Queensland in 2019, and is currently working toward the Ph.D. degree at the University of Queensland. His research interests include multimedia retrieval, machine learning and computer vision.


78-631; Data Science MM lab