Reducing bias to source samples for unsupervised domain adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised Domain Adaptation (UDA) aims to leverage labeled source data for unlabeled target data.
• Existing UDA methods often focus on domain alignment, assuming classifier generalization.
• This approach can fail when source and target domains have low similarity, leading to classifier bias.

Purpose of the Study:

• To propose a novel approach, RBDA, for unsupervised domain adaptation.
• To address the issue of classifier bias towards source domain features in UDA.
• To improve prediction accuracy on target domain data.

Main Methods:

• RBDA jointly matches domain distributions and reduces classifier bias to source samples.
• It utilizes adversarial networks conditioned with cross-covariance for distribution matching.
• Key mechanisms include a mean teacher model, regularization, and improved cross-entropy loss.

Main Results:

• RBDA achieves state-of-the-art results on several open benchmarks.
• The method demonstrates effectiveness in unsupervised domain adaptation scenarios.
• Experiments confirm the reduction of classifier bias to source samples.

Conclusions:

• RBDA offers an effective solution for UDA by tackling classifier bias.
• The proposed mechanisms successfully align domain distributions and enhance target predictions.
• RBDA advances the field of unsupervised domain adaptation, particularly for dissimilar domains.