Unsupervised Multi-Target Domain Adaptation: An Information Theoretic Approach.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Unsupervised domain adaptation (uDA) typically handles single source-single target domain settings.
• Real-world scenarios often involve adapting to multiple similar target domains simultaneously.
• Existing pairwise uDA methods are suboptimal for multi-target adaptation due to failure in leveraging shared domain information.

Purpose of the Study:

• To propose a novel information-theoretic approach for unsupervised domain adaptation (uDA) in a multi-target domain setting.
• To develop a model that finds a shared latent space across multiple domains while disentangling domain-specific private factors.
• To enable simultaneous adaptation from one labeled source domain to multiple unlabeled target domains.

Main Methods:

• An information-theoretic framework is employed to disentangle shared latent representations from domain-specific private factors.
• The model establishes a strong link between latent representations and observed data.
• An efficient optimization algorithm is developed for simultaneous multi-target adaptation.

Main Results:

• The proposed method successfully adapts a single source domain to multiple target domains simultaneously.
• The approach demonstrates superior performance compared to several popular uDA methods.
• Experiments conducted on three challenging public datasets validate the effectiveness of the proposed model.

Conclusions:

• The developed information-theoretic approach offers an effective solution for multi-target unsupervised domain adaptation.
• Disentangling shared and private information is crucial for leveraging multi-domain data.
• The method provides a significant advancement over traditional pairwise adaptation techniques.