Cycle Label-Consistent Networks for Unsupervised Domain Adaptation

TL;DR

This paper introduces Cycle Label-Consistent Networks (CLCN), a novel unsupervised domain adaptation method that uses cycle consistency of classification labels to improve feature discrimination across domains with minimal computational cost.

Contribution

The paper proposes a new cycle label-consistent loss for domain adaptation that enhances class discrimination and reduces reliance on pseudo-label accuracy, outperforming previous methods.

Findings

Improves accuracy by 9.4% on Office-31

Enhances feature discrimination across domains

Reduces negative impact of false pseudo-labels

Abstract

Domain adaptation aims to leverage a labeled source domain to learn a classifier for the unlabeled target domain with a different distribution. Previous methods mostly match the distribution between two domains by global or class alignment. However, global alignment methods cannot achieve a fine-grained class-to-class overlap; class alignment methods supervised by pseudo-labels cannot guarantee their reliability. In this paper, we propose a simple yet efficient domain adaptation method, i.e. Cycle Label-Consistent Network (CLCN), by exploiting the cycle consistency of classification label, which applies dual cross-domain nearest centroid classification procedures to generate a reliable self-supervised signal for the discrimination in the target domain. The cycle label-consistent loss reinforces the consistency between ground-truth labels and pseudo-labels of source samples leading to statistically similar latent representations between source and target domains. This new loss can easily be added to any existing classification network with almost no computational overhead. We demonstrate the effectiveness of our approach on MNIST-USPS-SVHN, Office-31, Office-Home and Image CLEF-DA benchmarks. Results validate that the proposed method can alleviate the negative influence of falsely-labeled samples and learn more discriminative features, leading to the absolute improvement over source-only model by 9.4% on Office-31 and 6.3% on Image CLEF-DA.