Thursday 10 April 2025
A team of researchers has made significant strides in developing a new method for unsupervised domain adaptation, a crucial challenge in machine learning. The approach, known as One-Shot Federated Unsupervised Domain Adaptation (UDA), allows models to adapt to new domains without requiring any additional labeled data.
Traditionally, UDA involves collecting and labeling data from multiple sources, which can be time-consuming and costly. However, with the rise of large-scale datasets and advancements in deep learning, researchers have been working on developing methods that can adapt to new domains using only unlabeled data.
The key innovation behind One-Shot Federated UDA is its ability to learn from a single example from each domain. This is achieved through a novel approach called Scaled Entropy Attention (SEA), which assigns weights to source models based on their uncertainty in predicting the target domain.
Uncertainty, in this context, refers to the model’s confidence in its predictions. When a model is uncertain, it means that it has difficulty distinguishing between different classes or making accurate predictions. SEA takes advantage of this uncertainty by giving more weight to models that are confident in their predictions and less weight to those that are uncertain.
The second component of One-Shot Federated UDA is Multi-Source Pseudo Labeling (MSPL), which generates pseudo labels for the target domain using a combination of source models. These pseudo labels are then used to train the global model, allowing it to adapt to the new domain without requiring any additional labeled data.
The results of this approach have been impressive, with significant improvements in performance on several benchmark datasets. In one experiment, the global model was able to achieve an accuracy of 83% on a target domain, compared to just 48% using traditional UDA methods.
One of the key advantages of One-Shot Federated UDA is its ability to adapt to new domains without requiring any additional data collection or labeling. This makes it particularly useful for applications where labeled data is scarce or expensive to collect.
Another advantage is its scalability, as it can be applied to large-scale datasets and complex models. The approach has been tested on several benchmark datasets, including OfficeHome and DomainNet, with impressive results.
In addition to its practical applications, One-Shot Federated UDA also has implications for our understanding of machine learning and deep learning. It highlights the importance of uncertainty in model predictions and demonstrates a new way of leveraging this uncertainty to improve performance.
Cite this article: “Unleashing Federated Learning: One-Shot Domain Adaptation via Scaled Entropy Attention and Multi-Source Pseudo Labeling”, The Science Archive, 2025.
Machine Learning, Unsupervised Domain Adaptation, One-Shot Federated Uda, Scaled Entropy Attention, Multi-Source Pseudo Labeling, Deep Learning, Uncertainty, Model Predictions, Scalability, Benchmark Datasets
