Joint Feature and Labeling Function Adaptation for Unsupervised Domain Adaptation

Abstract

Unsupervised domain adaptation aims to transfer knowledge from a labeled source domain to an unlabeled target domain. Although having achieved remarkable progress, most existing methods only focus on learning domain-invariant features and achieving a small source error. They ignore the discrepancy between labeling functions which will also cause discrepancy across domains. Inspired by this observation, we propose a novel method to simultaneously perform feature adaptation and labeling function adaptation. Specifically, for the feature adaptation, a domain discriminator is trained to reduce the discrepancy between feature distributions across domains. For the labeling function adaptation, we introduce a target predictor and a predictor discriminator. The target predictor is trained on target samples with pseudo-labels. The predictor discriminator is a novel component and is trained to distinguish whether the prediction output is from the source or the target predictor while the feature extractor and the label predictors try to confuse the predictor discriminator in an adversarial manner. Additionally, the intrinsic characteristics of the target domain are expected to be exploited thanks to the task-specific training. Comprehensive experiments are conducted and results validate the effectiveness of labeling function adaptation and demonstrate that our approach outperforms state-of-the-art methods.