NGSL

NGSL performs unsupervised domain adaptation by using nonlinear guide subspace learning (GSL) to construct invariant, discriminative subspaces that align non-independent and identically distributed (non-i.i.d.) source and target data.

Key Features:

• Invariant Subspace Learning: Constructs an invariant, discriminative, domain-agnostic subspace through guided refinement across two training stages.
• Subspace-Guided Term: Reduces domain discrepancy by aligning source data closer to the target subspace.
• Data-Guided Term: Uses coupled projections to map both domains into a unified subspace with low-rank coefficient matrices that preserve global data structure.
• Label-Guided Term: Leverages source labels and pseudo-target labels with a label relaxation matrix to enhance discrimination and robustness to label noise.
• Two-Stage Progressive Training Strategy: Employs a teacher-student feedback loop to iteratively refine discriminative domain-agnostic subspaces.
• Nonlinear Domain Adaptation: Extends the GSL framework with kernel embedding techniques to accommodate nonlinear domain shifts.

Scientific Applications:

• Unsupervised Domain Adaptation for non-i.i.d. data: Adapts labeled source-domain data to unlabeled target-domain data in machine learning tasks with distributional shifts.
• Cross-Domain Visual Recognition: Applies to cross-domain visual benchmark datasets for recognition and transfer learning evaluations.
• Bioinformatics Feature Extraction and Adaptation: Supports scenarios in bioinformatics that require robust, domain-invariant feature learning and adaptation.

Methodology:

Learn a unified invariant subspace via a three-tier guidance mechanism (subspace-, data-, and label-guided terms) using coupled projections and low-rank coefficient matrices; incorporate pseudo-target labels with a label relaxation matrix; employ a two-stage teacher-student progressive training loop; and apply kernel embedding techniques for nonlinear adaptation.