Knowledge Exchange Between Domain-Adversarial and Private Networks Improves Open Set Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open Set Domain Adaptation (OSDA) aims to generalize models to unseen domains while identifying unknown classes.
• Leveraging both domain-invariant and target-specific features is crucial for effective OSDA.
• Existing methods may not fully exploit the complementary nature of these features.

Purpose of the Study:

• To propose a novel Knowledge Exchange (KnowEx) model for Open Set Domain Adaptation.
• To enhance the discrimination between known and unknown classes in target domains.
• To improve the generalization capabilities of models across different data domains.

Main Methods:

• The KnowEx model jointly trains a Domain-Adversarial Network (DAdvNet) for domain-invariant features and a Private Network (PrivNet) for target-specific features.
• Adversarial perturbation is used to regularize PrivNet, enhancing complementarity with DAdvNet.
• An adaptation layer is incorporated into DAdvNet to address PrivNet's unreliability, fostering mutual reinforcement.

Main Results:

• The proposed KnowEx model demonstrates effectiveness in Open Set Domain Adaptation.
• Experimental results on standard benchmarks validate the superiority of the KnowEx approach.
• The joint training strategy mutually reinforces DAdvNet and PrivNet for improved performance.

Conclusions:

• The KnowEx model offers a robust framework for Open Set Domain Adaptation by effectively integrating domain-invariant and target-specific learning.
• The proposed method achieves superior performance compared to existing approaches on multiple benchmarks.
• KnowEx provides a promising direction for future research in domain adaptation and open set recognition.