PhD Proposal: Practical Robust Learning under Domain Shifts

Talk
Luyu Yang
Time: 
12.03.2021 10:00 to 12:00
Location: 

IRB 4105

The data we create is shifting rapidly. Despite the domain shifts among the images, we as humans still recognize the image. However, these shifts are a much bigger challenge for machines. The fundamental question is: how can we make machines as adaptive as humans? During my PhD, I have worked towards addressing this question through advances in the study of robust learning under domain shifts via domain adaptation.To enable real systems with demonstrated robustness, the study of domain adaptation needs to move from ideals to realities. In current domain adaptation research, these ideals fall into two categories: i) The trained model can generate invariant representations that work well on both the source and target domains. ii) The domain shift from the source to the target domain can be accurately modeled and then adapted. These ideals are generally not consistent with reality. First, in real scenarios, all resources are under a fixed budget. The model size cannot expand infinitely to handle the complexity of multiple domains. Second, there will not be datasets perfectly sliced into each domain annotated for alignment. Third, the domain shift often changes with time and therefore cannot be modeled only once. These real-world challenges are more than simple restrictions on the existing methods, and call for entirely new designs. In my dissertation research I propose to address these limitations from three aspects: First, domain adaptation should be time-sensitive; Second, true domain labels are hard to obtain; Finally, we need a new perspective to understand the robustness of domain adaptation.Examining Committee:

Chair:Department Representative:Members:

Dr. Abhinav Shrivastava Dr. Larry S. Davis Dr. David JacobsDr. Judy Hoffman (Georgia Tech)