25 Sep 2019 (modified: 24 Dec 2019)ICLR 2020 Conference Blind SubmissionReaders: Everyone
  • Original Pdf: pdf
  • TL;DR: We propose a reliable conditional adversarial learning scheme along with a simple, generic yet effective framework for UDA tasks.
  • Abstract: This paper presents a generic framework to tackle the crucial class mismatch problem in unsupervised domain adaptation (UDA) for multi-class distributions. Previous adversarial learning methods condition domain alignment only on pseudo labels, but noisy and inaccurate pseudo labels may perturb the multi-class distribution embedded in probabilistic predictions, hence bringing insufficient alleviation to the latent mismatch problem. Compared with pseudo labels, class prototypes are more accurate and reliable since they summarize over all the instances and are able to represent the inherent semantic distribution shared across domains. Therefore, we propose a novel Prototype-Assisted Adversarial Learning (PAAL) scheme, which incorporates instance probabilistic predictions and class prototypes together to provide reliable indicators for adversarial domain alignment. With the PAAL scheme, we align both the instance feature representations and class prototype representations to alleviate the mismatch among semantically different classes. Also, we exploit the class prototypes as proxy to minimize the within-class variance in the target domain to mitigate the mismatch among semantically similar classes. With these novelties, we constitute a Prototype-Assisted Conditional Domain Adaptation (PACDA) framework which well tackles the class mismatch problem. We demonstrate the good performance and generalization ability of the PAAL scheme and also PACDA framework on two UDA tasks, i.e., object recognition (Office-Home,ImageCLEF-DA, andOffice) and synthetic-to-real semantic segmentation (GTA5→CityscapesandSynthia→Cityscapes).
  • Keywords: Domain Adaptation, Transfer Learning, Adversarial Learning
8 Replies