Transfer Independently Together: A Generalized Framework for Domain Adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Unsupervised heterogeneous domain adaptation is crucial for real-world applications but remains underexplored.
• Existing methods often have limitations, such as requiring labeled target data or being restricted to specific scenarios.

Purpose of the Study:

• To address limitations in generalized unsupervised heterogeneous domain adaptation.
• To propose a novel framework, Transfer Independently Together (TIT), for effective domain alignment.

Main Methods:

• TIT learns independent transformations for each domain, mapping data to a shared latent space for alignment.
• A joint optimization framework unifies these transformations.
• A novel graph-optimization-based landmark selection algorithm reweights samples based on geometric relationships, avoiding complex matrix operations.

Main Results:

• The proposed landmark selection uses efficient integer arithmetic, outperforming existing float-point methods.
• TIT optimizes objectives via dimensionality reduction, applicable to arbitrary sample dimensions.
• Extensive evaluations demonstrate TIT's superiority on image classification, text categorization, and text-to-image recognition tasks.

Conclusions:

• TIT offers a robust and efficient solution for generalized unsupervised heterogeneous domain adaptation.
• The framework's ability to work without labeled target data significantly enhances its practical utility.
• TIT achieves state-of-the-art performance across diverse benchmarks and large-scale datasets.