Unsupervised Adaptation Across Domain Shifts by Generating Intermediate Data Representations.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Unconstrained data acquisition necessitates handling distribution shifts between training and testing datasets.
• Domain adaptation, particularly unsupervised domain adaptation, is crucial for real-world applications.

Purpose of the Study:

• To develop a robust, data-driven domain adaptation technique for unsupervised learning scenarios.
• To effectively bridge the domain gap using intermediate data representations.

Main Methods:

• A two-stage approach generating intermediate cross-domain representations via geodesic paths on Grassmann manifolds.
• Incorporation of non-linear domain representations (Reproducing Kernel Hilbert Space, Laplacian Eigenmaps).
• Exploration of semi-supervised and multi-domain adaptation settings, supplemented by reference domains and multi-class boosting.

Main Results:

• The proposed method generates informative intermediate representations to address domain shift.
• Competitive performance achieved on object recognition tasks using the Office and Bing datasets.
• Demonstrated effectiveness in unsupervised, semi-supervised, and multi-domain adaptation scenarios.

Conclusions:

• The developed domain adaptation technique offers a promising solution for handling data distribution shifts.
• The method's flexibility and performance make it suitable for various object recognition applications.