Deep discriminative causal domain generalization

Abstract

Domain generalization aims to generalize knowledge learned from multi-domains sources to a target domain whose statistical distribution is unknown. The mainstream approach to domain generalization involves learning domain-invariant features by aligning them across the domain distribution. However, these methods usually neglect the causal relationship between the instances and class labels; thus, the domain generalization model may learn the features that are not causally related to the label predictions and lead to potential misclassification. To address this problem, we propose a deep discriminative causal domain generalization (DDCDG), which can effectively learn causal feature representations. Specifically, we first imitate various possible intervention styles by applying an enriched data augmentation technique while preserving causal features relevant to the prediction task. We then designed a new regularization term for causal feature disentanglement, aiming to separate the causal and non-causal features. Additionally, we introduce centre alignment to approximate the representations of cross-domain homogeneous objects and enhance their discriminative ability. Extensive cross-domain visual recognition experiments were conducted, and the results show that our method can achieve a significant improvement in classification accuracy when compared with existing state-of-the-art methods on the Domainbed benchmark.