MEDIC meta-learning balances domains and classes for open set generalization

A team led by Xiran Wang at Nanjing University has introduced MEDIC, a dualistic meta-learning framework that tackles open set domain generalization—a harder variant of domain generalization where target domains contain both unseen distributions and unseen classes. Traditional domain generalization assumes label sets match between source and target, but real-world deployments often encounter novel categories. A naive fix using one-vs-all classifiers suffers from severe positive-negative imbalance, biasing decision boundaries toward over-rejecting out-of-distribution data, even for known classes.

MEDIC solves this by jointly optimizing gradients from two complementary meta-task splits: inter-domain and inter-class. The inter-domain split forces the model to generalize across environmental shifts, while the inter-class split learns to separate known from unknown categories. By aligning gradients from both splits implicitly, the model finds a Pareto-optimal decision boundary that balances domain invariance and class separability. Experimental results show MEDIC significantly surpasses prior state-of-the-art on multiple open set benchmarks (e.g., Office-Home, PACS), while maintaining competitive accuracy on standard closed-set domain generalization metrics.