Unsupervised and Semi-Supervised Robust Spherical Space Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Adversarial domain adaptation effectively learns domain-invariant features through adversarial training.
• Existing methods often struggle with leveraging target domain data, especially in semi-supervised scenarios.

Purpose of the Study:

• To propose a novel adversarial domain adaptation approach in a spherical feature space.
• To enhance feature learning for both unsupervised and semi-supervised domain adaptation.
• To improve the robust utilization of pseudo-labels and reduce intra-domain discrepancy.

Main Methods:

• Introduced a spherical feature space with spherical classifiers and discriminators.
• Developed a spherical robust pseudo-label loss using a Gaussian-uniform mixture model.
• Proposed a reweighted adversarial training strategy for semi-supervised domain adaptation.

Main Results:

• The proposed method achieves competitive or superior performance on object, digit, and face recognition benchmarks.
• Ablation studies validate the effectiveness of individual components: spherical classifier, discriminator, pseudo-label loss, and reweighted training.

Conclusions:

• The novel spherical adversarial domain adaptation approach is effective for both unsupervised and semi-supervised settings.
• The method demonstrates robust feature learning and improved performance across various recognition tasks.