Discriminative and Geometry-Aware Unsupervised Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation (DA) seeks to generalize models despite data distribution mismatches.
• Current DA methods primarily focus on aligning source and target data distributions.
• A theoretical analysis suggests a need for enhanced DA strategies beyond simple distribution alignment.

Purpose of the Study:

• To propose effective domain adaptation methods for classification tasks.
• To develop DA approaches that ensure feature alignment, class discriminability, and geometric structure awareness.
• To introduce Close Yet Discriminative DA (CDDA), Geometry-Aware DA (GA-DA), and Discriminative and GA-DA (DGA-DA) models.

Main Methods:

• Searching for a shared feature subspace that aligns distributions and separates classes.
• Incorporating the geometric structure of data manifolds into the adaptation process.
• Developing CDDA, GA-DA, and DGA-DA models to implement these principles.

Main Results:

• The proposed DGA-DA method consistently outperforms state-of-the-art DA techniques.
• Effectiveness demonstrated across 49 image classification DA tasks on eight benchmarks.
• In-depth analysis quantifies the contribution of each model component and provides data visualization insights.

Conclusions:

• Effective DA for classification requires both distribution alignment and discriminative feature learning.
• Accounting for data manifold geometry is crucial for accurate target domain label inference.
• The proposed DGA-DA method offers a robust and superior approach to domain adaptation in image classification.