Universal Domain Adaptation

Universal Domain Adaptation (UniDA) tackles the challenge of transferring knowledge from a labeled source dataset to an unlabeled target dataset when the datasets have differing label sets (category shift) and data distributions (domain shift). Current research focuses on developing methods that effectively identify and handle both "known" (shared) and "unknown" (target-specific) classes, often employing techniques like optimal transport, clustering algorithms (e.g., global-local clustering), and contrastive learning within various model architectures, including convolutional neural networks and vision transformers. UniDA's significance lies in its potential to improve the robustness and generalizability of machine learning models across diverse real-world applications, particularly in scenarios with limited labeled target data or evolving data distributions, such as in remote sensing and medical image analysis.