New Meta-Learning Boosts AI Generalization in Open Set Scenarios

Traditional domain generalization methods struggle when target domains introduce entirely new, unseen classes. Existing approaches to open set domain generalization often create imbalanced classifiers, leading to over-rejection of valid data from known classes in new domains. This imbalance arises because they treat unknown classes as a single negative category, skewing the decision boundary. A new meta-learning strategy, called MEDIC (dualistic MEta-learning with joint DomaIn-Class matching), addresses this by simultaneously considering inter-domain and inter-class task splits. It implicitly matches gradients to find optimal decision boundaries that are balanced for both different domains and different classes. Experimental results demonstrate that MEDIC not only surpasses previous methods in open set scenarios, where new classes appear, but also maintains competitive performance in traditional closed set generalization tasks. This indicates a more robust and adaptable AI model for real-world applications with evolving data distributions.