Active Learning for Unsupervised Domain Adaptation

Active learning methods have been explored to improve UDA by actively annotating a small subset of informative target domain samples. This chapter introduces two novel techniques to address key limitations of existing active domain adaptation (ADA) methods: estimating target representativeness without source data access and probabilistic uncertainty estimation. First, an energy-based criterion is proposed for selecting representative target samples without requiring source data, enabling application to source-free ADA. Second, a diffusion-based adversarial probabilistic model is presented to capture predictive uncertainty beyond deterministic point estimates, which leverages diffusion models for distributional classification. Experiments on benchmark datasets demonstrate the efficacy of both techniques, outperforming prior ADA methods. The probabilistic approach also naturally extends to source-free ADA. The proposed solutions provide more efficient and reliable uncertainty estimation for active domain adaptation. Experiments on three datasets demonstrate the efficacy of both techniques for tackling the challenges in ADA, including the more challenging source-free setting.