关键字：

摘要：Single Domain Generalization (SDG) is a realistically challenging casein domain generalization, only one available source domain and is committed to generalizing the knowledge acquired from domain to unseen target domains. Due to the lack of diversity information, data augmentation the data distribution is the mainstream method for SDG. Most low-level data augmentation methods model large domain shifts and simulated domains are fragile, which prevents the model from thoroughly improving the generalization ability. We bridge SDG to causal concepts, novelly analyzing the reason limited generated domain shifts restrict the improvement of generalization ability from the view backdoor path, and propose a causally stable framework CaRGI (Causal Semantic Representation Learning Generative Intervention). Firstly, we construct the inclusive causal directed acyclic graph and utilize analysis to gradually explore the relationship between the generated domain shift and generalization We regard domain expansion as the causal intervention and secondly propose the joint generative intervention module with dynamic to enlarge the domain shift with semantic consistency, which is dedicated to eliminating spurious confounding effects by blocking the backdoor path. Thirdly, counterfactual inference is implement causal semantic representation learning. In this process, the min-max game in the latent plays between the generative intervention module and the prediction module, which engage in alternating training that benefit each other in the spiral development. Specially, we innovatively perform maximization and minimization operations in shallow and deep layers respectively. CaRGI can finally learn causal representations and improve the stable generalization ability. Extensive experimental results on several used datasets verify the feasibility and effectiveness of the proposed method.

卷号：173

期号：

是否译文：否