Deep Subdomain Adaptation Network for Image Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation is crucial for machine learning when labeled data is unavailable in the target domain.
• Existing deep domain adaptation methods often align global distributions, neglecting crucial subdomain relationships and fine-grained information.
• Subdomain adaptation methods are emerging but often rely on slow, adversarial approaches.

Purpose of the Study:

• To develop a novel deep subdomain adaptation network (DSAN) for efficient and effective transfer learning.
• To address the limitations of existing methods by focusing on aligning relevant subdomain distributions.
• To provide a simple, fast-converging alternative to adversarial subdomain adaptation techniques.

Main Methods:

• Proposes the Deep Subdomain Adaptation Network (DSAN).
• Employs Local Maximum Mean Discrepancy (LMMD) to align domain-specific layer activations across relevant subdomains.
• Integrates LMMD loss into feedforward network models, trainable via standard backpropagation without adversarial training.

Main Results:

• DSAN demonstrates significant improvements in transfer learning performance.
• The method achieves remarkable results on object recognition and digit classification tasks.
• DSAN exhibits fast convergence due to its non-adversarial nature.

Conclusions:

• DSAN offers a simple yet effective approach for deep subdomain adaptation.
• The LMMD-based method provides efficient training and fast convergence.
• DSAN successfully captures fine-grained information by aligning relevant subdomains, outperforming previous methods.