A Causal Intervention Method for Domain Generalization with a Self-Supervised Auxiliary Task

Qinkang Gong, Yan Pan, Hanjiang Lai, Jian Yin

Abstract: Domain generalization tackles the challenge of domain shifts by learning a model from diverse source domains that can effectively generalize to unseen target domains. This paper explores domain generalization on image data with auxiliary learning task(s), which leverages auxiliary task(s) to extract transferable features and boost the performance of the primary domain generalization task on unseen domains. Causal intervention provides an attractive strategy to tackle domain shifts and learn causal dependencies in domain generalization tasks. However, most of the existing causal intervention methods are tailored for single-task learning. Causal intervention on multiple tasks (e.g., the primary and auxiliary tasks in domain generalization) remains under-explored. In this paper, we propose CI-DGA, a novel causal intervention method for domain generalization on image data with self-supervised auxiliary task(s). In CI-DGA, we employ a hidden confounder to model the data distribution of a specific domain, wherein this distribution changing with domain shifts. Theoretically, we show that the negative effects by this confounder can be eliminated by causal intervention, and the causal relations between the input images and the labels in both the primary and auxiliary tasks are identifiable. Additionally, we develop a deep architecture to implement the causal inference models, in which we provide an approximate strategy to reduce the computational cost and avoid simultaneous sampling operations on multiple variables. Comprehensive experimental results on three widely-used benchmark datasets show that the proposed CI-DGA has superior performance against state-of-the-art baselines for domain generalization on images.

External IDs:dblp:journals/ijcv/GongPLY25