Reserve to Adapt: Mining Inter-Class Relations for Open-Set Domain Adaptation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Open-Set Domain Adaptation (OSDA) addresses model adaptation challenges with unknown classes in target domains.
• Existing OSDA methods often struggle with negative transfer from unknown classes.
• A key challenge is effectively handling unknown classes without compromising known class performance.

Purpose of the Study:

• To propose a novel OSDA method that reserves embedding spaces for unknown classes.
• To leverage source domain inter-class relationships for improved target domain adaptation.
• To mitigate negative transfer caused by unknown classes in OSDA.

Main Methods:

• Reserving in-between embedding spaces for unknown classes instead of treating them as a single entity.
• Tightening known-class representations and enlarging inter-class margins in the source domain.
• Learning soft-label prototypes in the source domain to discriminate known and unknown samples in the target domain.
• Iteratively applying these steps to mutually enhance space reservation and unknown sample discrimination.

Main Results:

• Achieved state-of-the-art results on standard OSDA datasets: Office-31, Office-Home, VisDA, and ImageCLEF.
• Demonstrated the effectiveness of reserving embedding spaces for unknown classes.
• Provided analysis to understand the proposed method's behavior and benefits.

Conclusions:

• The proposed method effectively reserves space for unknown classes by utilizing source domain knowledge.
• This approach improves model adaptation in open-set scenarios by avoiding negative transfer.
• The findings offer a new perspective on handling unknown classes in domain adaptation.