Incremental Confidence Sampling with Optimal Transport for Domain Adaptation.

Area of Science:

• Machine Learning
• Statistical Learning Theory

Background:

• Domain adaptation addresses data distribution shifts between source and target domains.
• Unsupervised domain adaptation lacks labeled data in the target domain, posing a significant challenge.

Purpose of the Study:

• To present OTP-DA, an incremental approach for unsupervised domain adaptation.
• To develop a method that learns domain-invariant and well-separated joint subspaces.

Main Methods:

• Utilizes Linear Discriminant Analysis (LDA) to learn joint subspaces.
• Employs a selective label propagation technique based on optimal transport (OTP) to generate pseudo-labels for target data.
• Implements a self-training mechanism facilitated by pseudo-labels within latent subspaces.

Main Results:

• OTP-DA demonstrates promising efficacy and robustness in visual domain adaptation tasks.
• The proposed method shows favorable performance compared to state-of-the-art approaches.
• Theoretical analysis provides conditions for efficient unsupervised domain adaptation.

Conclusions:

• OTP-DA effectively overcomes the lack of target domain labels in unsupervised domain adaptation.
• The integration of optimal transport and self-training offers a robust solution for domain shift problems.
• The approach is validated through extensive experimentation on visual domain adaptation benchmarks.