Addressing Domain Shift in an Open World

The School of EECS is hosting the following seminar:

Addressing Domain Shift in an Open World

Speaker: Mr Zijian Wang
Host: Prof Helen Huang

Abstract: The success of the deep learning algorithms builds upon the assumption that the test data are independently identically distributed (i.i.d) with the training data. However, in practice, the i.i.d assumption is commonly violated by many factors, such as the change of illumination, weather, and environment, which causes significant performance degradation on the test data.  As a remedy to the performance degradation problem brought by domain shift, Domain adaptation (DA) and domain generalization (DG) aims to apply the previously learned knowledge to a different but related task are two representative topics under transfer learning.

In this thesis, we aim to address the more challenging heterogeneous domain adaptation, where the source and target domains have different input spaces or label spaces. Compared with DA, DG is recognized as a more realistic setting, since no target samples are required in the training stage. In contrast to the conventional DG that strictly requires the availability of multiple source domains, this thesis considers a more general yet challenging scenario, namely Single Domain Generalization (SDG), where only one source domain is available for training. To conclude, we discuss open questions and future work of improving the transferability and generalization power of deep learning models and identifying the limitations of our proposed approaches.

Speaker bio: Mr. Zijian Wang is a final year PhD student under the supervision of Prof. Helen Huang. His PhD thesis is mainly on domain adaptation and generalization in computer vision. His research outcomes have been published in ICCV, ICML, ACMMM, TMM, etc. Zijian has also been widely engaged in a number of cross-disciplinary research projects, spanning civil engineering and chemical engineering.