Decomposition-based transfer distance metric learning for image classification.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Distance Metric Learning (DML) is crucial for image analysis but requires extensive labeled data.
• Transfer learning addresses data scarcity by leveraging related source tasks, but differing data distributions pose challenges for existing DML methods.

Purpose of the Study:

• To develop a novel transfer learning approach for DML that overcomes domain shift issues.
• To enable robust metric learning for target tasks with limited side information by exploiting source task knowledge.

Main Methods:

• Proposes Decomposition-based Transfer DML (DTDML), representing the target metric as a sparse combination of base metrics derived from source task eigenvectors or random bases.
• Learns coefficients for base metrics rather than the target metric directly, reducing the number of learnable variables.

Main Results:

• DTDML demonstrates effectiveness in transfer metric learning, outperforming existing methods on handwritten digit/letter classification and natural image annotation.
• The method achieves more reliable solutions with limited side information and faster optimization due to learning fewer variables.

Conclusions:

• DTDML offers an effective solution for transfer DML by addressing domain shift through a decomposition-based approach.
• The proposed method provides a more efficient and reliable way to learn distance metrics in scenarios with limited labeled data.