Deep Ladder-Suppression Network for Unsupervised Domain Adaptation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) transfers knowledge from labeled source domains to unlabeled target domains.
• Existing UDA methods often struggle by adapting domain-specific image variations, hindering performance.
• This adaptation of domain-specific variations can undermine the effectiveness of learned features.

Purpose of the Study:

• To propose a novel module, the Deep Ladder-Suppression Network (DLSN), for enhanced UDA.
• To enable learning of cross-domain shared content by effectively suppressing domain-specific variations.
• To improve the performance of existing UDA frameworks by integrating DLSN.

Main Methods:

• The proposed DLSN is an autoencoder architecture with lateral connections from encoder to decoder.
• Domain-specific details are directly fed to the decoder for reconstruction, easing the encoder's task.
• This design allows the shared encoder to prioritize learning cross-domain shared content.

Main Results:

• Extensive experiments were conducted on four benchmark datasets: Digits, Office31, Office-Home, and VisDA-C.
• The DLSN module consistently and significantly improved the performance of various popular UDA frameworks.
• The method effectively suppresses domain-specific variations, allowing focus on shared content.

Conclusions:

• The Deep Ladder-Suppression Network (DLSN) is an effective module for unsupervised domain adaptation.
• DLSN enhances UDA by learning domain-invariant features through suppression of domain-specific variations.
• The proposed module offers a flexible and impactful addition to existing UDA approaches.