Gradual Domain Adaptation via Normalizing Flows.

Area of Science:

• Computer Science
• Machine Learning

Background:

• Standard domain adaptation methods struggle with significant differences between source and target data domains.
• Gradual domain adaptation uses intermediate domains but often fails with limited intermediate steps and large domain distances.

Purpose of the Study:

• To address the failure of gradual domain adaptation under challenging domain gaps.
• To propose a novel method using normalizing flows within unsupervised domain adaptation.

Main Methods:

• The proposed method employs normalizing flows to learn a transformation from target domain distributions to a Gaussian mixture distribution, utilizing the source domain.
• This approach maintains the unsupervised domain adaptation framework.

Main Results:

• Experiments on real-world datasets demonstrate the method's effectiveness in mitigating issues caused by large domain gaps.
• The proposed technique significantly improves classification performance compared to existing methods.

Conclusions:

• Normalizing flows offer a viable solution for gradual domain adaptation problems with large domain gaps.
• The method successfully enhances classification accuracy in challenging unsupervised domain adaptation scenarios.