analiza aiqbench
This research introduces EUDA, a parameter-efficient framework for unsupervised domain adaptation that freezes the DINOv2 backbone while updating only lightweight components. By combining cross-domain alignment losses with self-supervised vision transformers, the method significantly reduces trainable parameters without sacrificing performance on standard benchmarks like Office-Home and VisDA-2017. The approach addresses critical scalability issues in resource-constrained environments by eliminating the need for full fine-tuning of large models.
Efficient unsupervised domain adaptation via self-supervised vision transformer and synergistic cross-domain alignment
arXiv:2407.21311v2 Announce Type: replace-cross Abstract: Unsupervised domain adaptation (UDA) aims to mitigate domain shift, where the distribution of labeled source data differs from that of unlabeled target data. Despite recent advances, existing methods often rely on fine-tuning large backbone models, which leads to high computational cost and limits scalability in resource-constrained environments. This limitation highlights the need for parameter-efficient approaches that maintain strong performance with reduced training complexity. Self-supervised foundation models such as DINOv2 provide highly transferable representations and raise the question of whether effective domain adaptation can be achieved without full fine-tuning. To address this question, we propose Efficient Unsupervised Domain Adaptation (EUDA), a parameter-efficient framework that leverages a frozen DINOv2 backbone as a feature extractor and updates only a lightweight bottleneck and classification head. We also adopt a synergistic domain alignment loss (SDAL), which combines cross-entropy (CE) and maximum mean discrepancy (MMD) to promote both discriminative learning and cross-domain alignment. Experimental results on Office-Home, Office-31, VisDA-2017, and DomainNet demonstrate that EUDA achieves competitive performance across diverse domain complexities, while reducing the number of trainable parameters by 42 to 99.7%. These results show the suitability of the proposed method for resource-constrained and distributed environments.
Abstract:Unsupervised domain adaptation (UDA) aims to mitigate domain shift, where the distribution of labeled source data differs from that of unlabeled target data. Despite recent advances, existing methods often rely on fine-tuning large backbone models, which leads to high computational cost and limits scalability in resource-constrained environments. This limitation highlights the need for parameter-efficient approaches that maintain strong performance with reduced training complexity. Self-supervised foundation models such as DINOv2 provide highly transferable representations and raise the question of whether effective domain adaptation can be achieved without full fine-tuning. To address this question, we propose Efficient Unsupervised Domain Adaptation (EUDA), a parameter-efficient framework that leverages a frozen DINOv2 backbone as a feature extractor and updates only a lightweight bottleneck and classification head. We also adopt a synergistic domain alignment loss (SDAL), which combines cross-entropy (CE) and maximum mean discrepancy (MMD) to promote both discriminative learning and cross-domain alignment. Experimental results on Office-Home, Office-31, VisDA-2017, and DomainNet demonstrate that EUDA achieves competitive performance across diverse domain complexities, while reducing the number of trainable parameters by 42 to 99.7%. These results show the suitability of the proposed method for resource-constrained and distributed environments.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2407.21311 [cs.CV]
(or arXiv:2407.21311v2 [cs.CV] for this version)
Abedi, A., Wu, Q.M.J., Zhang, N. et al. Efficient unsupervised domain adaptation via self-supervised vision transformer and synergistic cross-domain alignment. Int. J. Mach. Learn. & Cyber. 17, 423 (2026)
https://doi.org/10.1007/s13042-026-03248-5 DOI(s) linking to related resources
Text integral sub licența deschisă a publicației · 245 cuvinte
Subiecte conexe
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Latent Space · 13 oct. 2026
Parameter-Efficient Adaptation of Pre-Trained Vision Foundation Models for Active and Passive Seismic Data Denoising
arXiv · 23 sept. 2026
HierDAMap: Towards Universal Domain Adaptive BEV Mapping via Hierarchical Perspective Priors
arXiv · 17 aug. 2026
WinoTS: Wavelet-based Self-Distillation for Time Series Models
arXiv · 1 oct. 2026