Deep Transfer Low-Rank Coding for Cross-Domain Learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Transfer learning leverages existing knowledge for new tasks, especially with limited data.
• Deep learning models face challenges with domain mismatch, particularly in higher layers.
• Existing methods struggle to effectively bridge domain gaps in deep transfer learning.

Purpose of the Study:

• To develop a novel deep transfer learning framework using low-rank coding.
• To address domain mismatch issues in deep convolutional neural networks.
• To enhance knowledge transfer for improved target domain recognition.

Main Methods:

• Developed a deep transfer low-rank coding method based on deep convolutional neural networks.
• Investigated multilayer low-rank coding at task-specific layers.
• Utilized shared common dictionaries and rank minimization for domain-invariant feature extraction.
• Integrated domain/classwise adaptation terms for semi-supervised optimization.

Main Results:

• The proposed method effectively bridges the domain gap by capturing domain-invariant knowledge.
• Rank minimization preserves global structures, promoting sample similarity across domains.
• Semi-supervised adaptation terms alleviate marginal and conditional disparities.
• Experimental results on visual domain adaptation benchmarks show significant performance improvement.

Conclusions:

• The novel deep transfer low-rank coding approach effectively enhances recognition performance in target domains.
• The method successfully mitigates domain mismatch issues in deep transfer learning.
• This approach offers a robust solution for visual domain adaptation tasks.