Semi-supervised Deep Domain Adaptation via Coupled Neural Networks.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation addresses limited labeled target data by leveraging source data.
• Multi-layer structures are used for discriminative feature learning to reduce domain discrepancy.
• Simultaneous deep structure and classifier learning for domain adaptation is under-explored.

Purpose of the Study:

• Propose a semi-supervised deep domain adaptation framework.
• Jointly learn a multi-layer feature extractor and a multi-class classifier.
• Improve feature transferability and alleviate domain divergence.

Main Methods:

• Develop a novel semi-supervised class-wise adaptation approach.
• Assign probabilistic labels to target samples to handle conditional distribution mismatch.
• Simultaneously train a multi-class classifier on labeled source and unlabeled target data.

Main Results:

• The proposed framework effectively reduces domain discrepancy.
• Enhanced feature transferability was observed.
• Experimental evaluations on standard benchmarks demonstrate superior performance.

Conclusions:

• The joint learning of deep structure and classifier is beneficial for domain adaptation.
• Semi-supervised class-wise adaptation effectively tackles distribution mismatch.
• The approach shows significant improvements in cross-domain tasks.