Enhancing Open-Set Domain Adaptation through Optimal Transport and Adversarial Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open-Set Domain Adaptation (OSDA) transfers knowledge from source to target domains where source classes are a subset of target classes.
• A key challenge in OSDA is aligning domains while separating known and unknown target classes.
• Existing methods often treat unknown classes as a single group, limiting discriminative information and decision boundary clarity.

Purpose of the Study:

• To propose a novel framework, Optimal Transport and Adversarial Learning (OTAL), to address limitations in current OSDA methods.
• To enhance the acquisition of discriminative information within the target domain.
• To improve the decision boundaries between known and unknown classes in OSDA.

Main Methods:

• Introduced Optimal Transport (OT) with a similarity matrix for feature-to-prototype mapping in clustering.
• Employed a three-way domain discriminator for aligning known sample distributions and constructing decision boundaries.
• Evaluated the framework on standard image classification datasets: Office-31, Office-Home, and VisDA-2017.

Main Results:

• The proposed OTAL framework demonstrated superior performance compared to existing state-of-the-art methods.
• Optimal Transport facilitated learning discriminative information and capturing the target domain's intrinsic structure.
• The three-way domain discriminator effectively aided in separating known and unknown classes.

Conclusions:

• OTAL effectively overcomes limitations of previous OSDA approaches by improving discriminative information acquisition and decision boundary clarity.
• The integration of Optimal Transport and adversarial learning offers a promising direction for advancing OSDA research.
• The framework shows significant potential for real-world applications requiring robust domain adaptation with unknown classes.