.png){kind=icon}
Return
Mixup-based maximum distribution difference selection strategy for domain generalization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127521.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Domain generalization is a branch of transfer learning problem where the source and target domains are distributed differently, and the goal is to generalize the model trained on the source domains to the unseen target domain. A promising solution is to apply mixup to generate new features, but existing methods typically directly select two samples from different domains for mixup. We find that this cross-domain selection method has the disadvantages of requiring domain labels and failing when the model has multiple mixups, etc. For this reason, we design a new sample selection algorithm in this paper to select pairs of samples with large differences in distribution for mixup. Further, we propose an indirect multi-sample mixup approach to explore the distribution of unseen target domains on a larger scale. To address the problem of overfitting caused by original data in the later stage of training, we design a formula to adjust the mix ratio as training proceeds, thus gradually increasing the mix ratio of the original data. Finally, extensive experiments on four public benchmark datasets, PACS, OfficeHome, VLCS, and TerraIncognita demonstrate that our method performs better compared to state-of-the-art methods for domain generalization.
Keywords:
Domain generalization
Transfer learning
Maximum Mean Discrepancy
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
Organization
No organization information available
Cited Papers
.png){kind=icon}
ShareSave
ShareSaveUnbiased Metric Learning: On the Utilization of Multiple Datasets and Web Images for Softening Bias
PREAI
ShareSaveShareSave
View more