Domain Adaptation Without Source Data

Domain adaptation assumes that samples from source and target domains are freely accessible during a training phase. However, such an assumption is rarely plausible in the real world possibly causes data privacy issues, especially when label of domain can be sensitive attribute as identifier. To avoid accessing could contain information, we introduce free (SFDA). Our key idea to leverage pretrained model progressively update self-learning manner. We observe with lower self-entropy measured by more likely classified correctly. From this, select reliable criterion define these class prototypes. then assign pseudolabels for every sample based on similarity score further propose point-to-set distance-based filtering, which does not require any tunable hyperparameters reduce uncertainty pseudolabeling process. Finally, train filtered regularization model. Surprisingly, without direct usage labeled samples, our SFDA outperforms conventional methods benchmark datasets. Impact Statement—This study addresses issue, unsupervised adaptation. Based privacy-preserving adaptation, various stakeholders, including enterprises government organizations, concern about issues their dataset. Furthermore, proposed data-free approach contribute creating positive social impact, large-scale Recently, since size across fields has been scaling up, it almost incapable individual researchers directly utilize large scale training. For this reason, new trend sharing models, e.g., EfficientNet BERT, led global huge amount resources rising up. viewpoint, thus enables people participate field specifically attributes.