Semantically Guided Gradient Matching for Open-Set Domain Generalization

Neural networks assume training and test data share the same distribution and label space, but real-world violations degrade performance when domain and category shifts occur simultaneously. Open Set Domain Generalization addresses recognizing unseen classes in unseen domains. Current approaches using one-vs-all classifiers suffer from biased decision boundaries due to class imbalance, while meta-learning methods ignore semantic relationships between classes. This paper proposes a semantically-guided gradient matching framework extending dualistic meta-learning with joint domain-class matching by incorporating a semantic encoder for modeling continuous class relationships. The key innovation is weighted gradient matching using semantic similarity to guide decision boundary formation. Experiments on PACS dataset show this approach outperforms previous methods in open set scenarios while maintaining competitive closed-set generalization, resulting in more balanced decision boundaries.