Domain Generalization by Joint-Product Distribution Alignment
作者：
Highlights：
• To address the domain difference in domain generalization, we propose to align multiple domains P1(x,y),…,Pn(x,y) via the alignment of two distributions: the joint distribution P(x,y,l) and the product distribution P(x,y)P(l), where the domain label l∈{1,⋯,n}.
• We analytically derive an explicit estimate of the Relative Chi-Square (RCS) divergence between P(x,y,l) and P(x,y)P(l), and minimize this estimate to align distributions in the neural transformation space.
• We demonstrate the effectiveness of our solution through conducting comprehensive experiments on several multi-domain image classification datasets.
摘要
•To address the domain difference in domain generalization, we propose to align multiple domains P1(x,y),…,Pn(x,y) via the alignment of two distributions: the joint distribution P(x,y,l) and the product distribution P(x,y)P(l), where the domain label l∈{1,⋯,n}.•We analytically derive an explicit estimate of the Relative Chi-Square (RCS) divergence between P(x,y,l) and P(x,y)P(l), and minimize this estimate to align distributions in the neural transformation space.•We demonstrate the effectiveness of our solution through conducting comprehensive experiments on several multi-domain image classification datasets.
论文关键词：Distribution alignment,Distribution divergence,Domain generalization,Feature transformation
论文评审过程：Received 2 April 2022, Revised 15 September 2022, Accepted 28 September 2022, Available online 2 October 2022, Version of Record 14 October 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.109086
