TSTELM: Two-Stage Transfer Extreme Learning Machine for Unsupervised Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Extreme Learning Machine (ELM) excels in classification and regression due to speed and generalization.
• ELM struggles with domain adaptation when training and testing data distributions differ.

Purpose of the Study:

• Propose a novel ELM variant, the Two-Stage Transfer Extreme Learning Machine (TSTELM), to address domain adaptation challenges.
• Enhance knowledge transfer capabilities of ELM for improved performance on target domains with different data distributions.

Main Methods:

• Employ Maximum Mean Discrepancy (MMD) at the statistical matching stage to reduce output layer distribution differences.
• Implement subspace alignment, target cross-domain mean approximation, and output weight approximation for parameter alignment and knowledge transfer.
• Integrate ELM parameters from both stages for final test sample prediction.

Main Results:

• TSTELM effectively enhances knowledge transfer ability in domain adaptation tasks.
• Experiments on four public datasets demonstrate higher accuracy compared to existing transfer and non-transfer classifiers.
• The proposed method shows robust performance in classification tasks across different domains.

Conclusions:

• TSTELM offers a significant improvement over standard ELM for domain adaptation.
• The two-stage approach successfully bridges the distribution gap between source and target domains.
• TSTELM provides a promising solution for machine learning applications requiring robust cross-domain generalization.