Title: Advancing Precision in Model Selection for Deep Unsupervised Domain Adaptation

Abstract: Deep unsupervised domain adaptation (Deep UDA) techniques effectively utilize abundant labeled data from a source domain to enhance performance on related, unlabeled target data. Nevertheless, the field faces a significant hurdle in algorithm comparison, primarily due to the lack of a standardized and precise model selection framework, which impedes further progress. Current model selection approaches for Deep UDA are often flawed, exhibiting issues such as high bias, restrictive applicability, instability, or controversy—particularly those that depend on labeled target data. To address these challenges, we introduce \textit{Deep Embedded Validation} (\textbf{DEV}). This method integrates adapted feature representations into the validation process to provide an unbiased estimate of target risk with controlled variance. Additionally, we employ control variates to further minimize this variance. The robustness of our approach is substantiated through both theoretical analysis and empirical evidence.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC