Robust Visual Knowledge Transfer via Extreme Learning Machine Based Domain Adaptation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Visual knowledge adaptation is crucial for robust visual categorization.
• Leveraging labeled source data with limited target data presents a significant challenge.
• Existing domain adaptation methods often struggle with effective knowledge transfer.

Purpose of the Study:

• To propose a novel Extreme Learning Machine (ELM) based Domain Adaptation (EDA) framework.
• To enable robust visual categorization by adapting classifiers across domains with minimal target labels.
• To enhance cross-domain learning by integrating unlabeled target data and manifold regularization.

Main Methods:

• Developed an EDA framework combining category transformation and ELM classification with random projection.
• Minimized the -norm of network output weights and learning error simultaneously.
• Integrated unlabeled target data as a fidelity term and employed manifold regularization with a Laplacian graph.

Main Results:

• The proposed EDA framework analytically determines and transfers network output weights.
• Experiments on benchmark datasets for video event and object recognition show superior performance.
• The MvEDA model, an extension considering multiple views, was also proposed and evaluated.

Conclusions:

• EDA provides a stable and effective solution for visual knowledge adaptation.
• The framework demonstrates significant improvements over existing cross-domain learning techniques.
• The method is beneficial for semi-supervised learning scenarios and adaptable to various base classifiers.