Meta Adversarial Weight for Unsupervised Domain Adaptation

Chang Liu, Lichen Wang, Yun Fu

Abstract: Despite great progress in supervised image recognition, a large performance drop is usually observed when deploying the model in the wild. Unsupervised domain adaptation (UDA) methods tackle the issue by aligning the source domain and the target domain. However, most existing adversarial based methods attempt to perform the alignment from a holistic view, ignoring the underlying class-level data structure in the target domain. As a result, the representations are distorted by adversarial alignment, leading to a negative transfer. Motivated by this issue, we first claim that this issue can be solved if there exists ‘optimal’ per-sample weights for adversarial alignment, and then devise a meta-learning framework to adaptively learn such adversarial weights. Specifically, we construct a meta-dataset with targetlike distribution as meta knowledge, and use it to guide the learning of the optimal adversarial weights via a meta-learner. By this means, our framework can adaptively adjust the weights of all training samples in adversarial training based on the feedback from meta dataset and thus achieve the categorical-wise domain alignment. We conduct sufficient ablation studies and experiments to show the effectiveness of our approach. Our method is generic to existing domain alignment based methods and could achieve consistently improvements over three UDA classification benchmarks.