Optimal Transport for Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Data Analytics

Background:

• Domain adaptation is crucial for robust machine learning models facing diverse data sources.
• Existing methods often struggle to generalize effectively across different observation systems.
• Domain-invariant representations offer a promising strategy for unified model training.

Purpose of the Study:

• To develop a novel domain adaptation technique for aligning data representations.
• To enhance model robustness and performance when applied to new domains.
• To exploit both labeled source data and unlabeled target data distributions.

Main Methods:

• A regularized unsupervised optimal transportation model is proposed for representation alignment.
• A transportation plan is learned to match probability density functions (PDFs) of source and target domains.
• Labeled source samples are constrained to remain close during the transportation process.

Main Results:

• The proposed method consistently outperforms state-of-the-art approaches on visual domain adaptation tasks.
• Experiments demonstrate improved performance on domain-invariant deep learning features.
• The approach shows adaptability to semi-supervised learning scenarios with limited target domain labels.

Conclusions:

• The optimal transportation-based domain adaptation method offers a powerful way to align data representations.
• This technique enhances model generalization and robustness across different domains.
• The method is effective and versatile, applicable to both unsupervised and semi-supervised learning settings.