Abstract:

Recent development of deep learning technologies benefits a variety of applications in computer vision. Among research topics in deep learning and computer vision, transfer learning particularly focuses on bridging information across data domains, so that feature representation or learning models observed in one (source) domain can be applied to another (target) domain of interest. In practice, since one might not be able to collect or annotate ground truth labels for the target-domain data, how to advance deep transfer learning techniques for solving this task would be very challenging. In this talk, I will cover our recent ICCV/AAAI/CVPR works on semantic segmentation, multi-label classification, image translation and representation disentanglement, which are all associated with learning from cross-domain data. I will discuss how we advance and extend existing deep learning models, and adapt such models for addressing the above tasks of interest.