Joint Domain Alignment and Adversarial Learning for Domain Generalization

Domain generalization aims to extract a classifier model from multiple observed source domains, and then can be applied to unseen target domains. The primary challenge in domain generalization lies in how to extract a domain-invariant representation. To tackle this challenge, we propose a multi-source domain generalization network called Joint Domain Alignment and Adversarial Learning (JDAAL), which learns a universal domain-invariant representation by aligning the feature distribution of multiple observed source domains based on multi-kernel maximum mean discrepancy. We adopt an optimal multi-kernel selection strategy that further enhances the effectiveness of embedding matching and approximates different distributions in the domain-invariant feature space. Additionally, we use an adversarial auto-encoder to bound the multi-kernel maximum mean discrepancy for rendering the feature distribution of all observed source domains more indistinguishable. In this way, the domain-invariant representation generated by JDAAL can improve the adaptability to unseen target domains. Extensive experiments on benchmark cross-domain datasets demonstrate the superiority of the proposed method.