F-MMD-DBA: Frobenius-norm Maximum Mean Discrepancy for domain bi-classifier adversarial

Abstract

Unsupervised domain bi-classifier adversarial approaches are promising methods to deal with domain shifts. Combined with metric learning, they significantly reduce ambiguous predictions. However, they have challenges in the leverage of global information, the relationships between subdomains, the consistency of feature representation, and the conflict due to the bi-classifier adversarial paradigm and usage of local information. Here, we propose Frobenius-norm Maximum Mean Discrepancy for Domain Bi-classifier Adversarial (F-MMD-DBA) to alleviate these by imposing and integrating global and local constraints. Maximum Mean Discrepancy, as the global constraint, can utilize global information. As the local constraint, Frobenius-norm focuses on the relationship of subdomains and avoids conflict. The integration of global and local constraints increases the consistency of feature representation. Ours has greatly improved the accuracy of digits classification and object recognition tasks. The code will be available at https://github.com/JackyDeepLearning/F-MMD-DBA.