BEIJING – Domain adaptation remains a hurdle in artificial intelligence, particularly when models trained in one area are applied to another.

Researchers at Southeast University and Lenovo Research said they have developed a new framework, Collective Domain Adversarial Learning (ColDA), to address shortcomings in conventional approaches.

Published in Frontiers of Computer Science, the study introduces a set-level adversarial method that leverages collective interactions among groups of samples rather than focusing on individual instances.

The framework employs a Dual-Set Domain Discriminator with attention mechanisms to capture both intra-domain and inter-domain relationships, improving feature alignment across domains.

Tests on benchmark datasets including VisDA-2017 and Office-Home showed ColDA outperforming existing state-of-the-art methods, particularly in noisy or ambiguous scenarios.

The researchers said future work will extend the approach to multi-domain and open-set adaptation to boost robustness and generalization.