Wasserstein Uncertainty Estimation for Adversarial Domain Matching.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Domain adaptation seeks to generalize models from a labeled source domain to an unlabeled target domain.
• Domain shift presents a significant challenge, hindering model performance on unseen data.

Purpose of the Study:

• To propose a novel domain adaptation technique by quantifying cross-domain transferability using domain prediction uncertainty.
• To develop a sample reweighting mechanism based on this uncertainty to mitigate domain shift.
• To enhance model generalization without requiring target domain fine-tuning.

Main Methods:

• Quantifying cross-domain transferability via domain prediction uncertainty using Wasserstein gradient flows.
• Implementing a sample reweighting strategy to adaptively exclude domain-specific samples during training.
• Evaluating the approach on benchmark datasets for both balanced and partial domain adaptation scenarios.

Main Results:

• The proposed reweighting mechanism effectively reduces domain shift.
• The method provides a meaningful curriculum for cross-domain transfer.
• Experimental results show improved performance in both balanced and partial domain adaptation settings.

Conclusions:

• Domain prediction uncertainty, quantified by Wasserstein gradient flows, is a viable metric for assessing cross-domain transferability.
• Adaptive sample reweighting based on this uncertainty is an effective strategy for domain adaptation.
• The proposed method enhances model generalization and achieves state-of-the-art results on benchmark datasets.