Robust Transfer Metric Learning for Image Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Metric learning is crucial for image analysis but typically requires extensive labeled data.
• Conventional methods assume similar data distributions between training and testing, which is often not the case.
• High labeling costs limit the practical application of supervised metric learning.

Purpose of the Study:

• To develop a robust transfer metric learning (RTML) framework for effective knowledge transfer from labeled source domains to unlabeled target domains.
• To mitigate domain shift in both sample and feature spaces.
• To enhance robustness against noisy data in metric learning.

Main Methods:

• RTML framework utilizes knowledge transfer to bridge distribution disparities between source and target domains.
• Domain-wise and class-wise adaptation schemes are employed in the sample space.
• A marginalized denoising approach with a low-rank constraint is used in the feature space.
• An explicit rank constraint regularizer replaces the NP-hard rank minimization problem.

Main Results:

• Experimental results on standard benchmarks demonstrate the effectiveness of the proposed RTML framework.
• RTML outperforms existing state-of-the-art transfer learning and metric learning algorithms.
• The framework successfully mitigates domain shift and handles noisy data.

Conclusions:

• The developed RTML framework provides a robust solution for metric learning with limited labeled data.
• Knowledge transfer across domains significantly improves performance in unlabeled target learning.
• The proposed methods for handling domain shift and noise are effective for real-world applications.