Non-Convex Transfer Subspace Learning via Embedded Distribution Alignment.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Transfer subspace learning is key for unsupervised domain adaptation.
• Existing methods often neglect cross-domain joint probability distribution alignment in subspaces.
• This limits the effectiveness of learned low-dimensional representations.

Purpose of the Study:

• To propose a novel non-convex transfer learning method, DATSL, for improved unsupervised domain adaptation.
• To address the limitations of existing methods in aligning cross-domain distributions.
• To enhance the discriminability of learned representations in shared embedding spaces.

Main Methods:

• Developed DATSL, a non-convex transfer learning method using embedded distribution alignment.
• Incorporated a non-convex regularizer minimizing singular values to approximate low-rank constraints.
• Introduced category-aware joint distribution alignment via label-informed covariance matching.
• Extended DATSL to GDATSL with manifold-preserving Laplacian regularization.
• Designed an efficient iterative optimization algorithm with proven convergence.

Main Results:

• DATSL effectively aligns joint distributions across domains.
• Category-aware alignment enhances representation discriminability and subspace discriminability.
• GDATSL preserves intrinsic data topology during knowledge transfer.
• Extensive experiments show superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed DATSL and GDATSL methods significantly advance unsupervised domain adaptation.
• Embedded distribution alignment and category-aware mechanisms are effective.
• The methods demonstrate robust performance across various datasets.