Propensity-driven Uncertainty Learning for Sample Exploration in Source-Free Active Domain Adaptation

TL;DR

Propensity-driven Uncertainty Learning (ProULearn) tackles SFADA by selecting informative target samples before training using a Homogeneity Propensity Estimation (HPE) and a correlation-based linkage among samples, thereby guiding robust pseudo-labeling without source data or continual human annotations. It introduces a central correlation loss ($L_{cc}$) and a weighted cross-entropy plus information maximization objective to produce compact class clusters and stable domain alignment. Across four SFADA benchmarks, ProULearn yields consistent improvements over state-of-the-art methods, demonstrating strong performance with only $5\%$ labeled target data and no access to source data. The framework's informative-sample strategy and correlation-based learning offer broad applicability to deep learning tasks requiring identification of representative samples and feature relationships.

Abstract

Source-free active domain adaptation (SFADA) addresses the challenge of adapting a pre-trained model to new domains without access to source data while minimizing the need for target domain annotations. This scenario is particularly relevant in real-world applications where data privacy, storage limitations, or labeling costs are significant concerns. Key challenges in SFADA include selecting the most informative samples from the target domain for labeling, effectively leveraging both labeled and unlabeled target data, and adapting the model without relying on source domain information. Additionally, existing methods often struggle with noisy or outlier samples and may require impractical progressive labeling during training. To effectively select more informative samples without frequently requesting human annotations, we propose the Propensity-driven Uncertainty Learning (ProULearn) framework. ProULearn utilizes a novel homogeneity propensity estimation mechanism combined with correlation index calculation to evaluate feature-level relationships. This approach enables the identification of representative and challenging samples while avoiding noisy outliers. Additionally, we develop a central correlation loss to refine pseudo-labels and create compact class distributions during adaptation. In this way, ProULearn effectively bridges the domain gap and maximizes adaptation performance. The principles of informative sample selection underlying ProULearn have broad implications beyond SFADA, offering benefits across various deep learning tasks where identifying key data points or features is crucial. Extensive experiments on four benchmark datasets demonstrate that ProULearn outperforms state-of-the-art methods in domain adaptation scenarios.

Figures (5)

• Figure 1: t-SNE plots show the selected sample location (red stars) and overall feature space before domain adaptation and after training. The experiments are based on the Office-31 dataset (A$\rightarrow$W) with 31 classes and 795 target domain samples.
• Figure 2: Overall ProULearn framework during target domain adaptation. The informative samples are selected before the training using HPE and correlation entropy. These active samples are fixed during the adaptation process. Meanwhile, the model and pseudo labels are refined during training.
• Figure 3: Principle of the homogeneity propensity estimation mechanism. An ensemble of trees is used to estimate sample homogeneity, where longer paths indicate more grouped samples. The mechanism selects representative grouped samples (green circles) rather than outliers (grey circles), guiding the model to focus on these samples during adaptation. This approach results in more compact and separable class distributions, as evidenced by the results shown in Figure \ref{['motivation']}.
• Figure 4: Ablation studies on hyper-parameters (\ref{['hyper_parameter']}) and central correlation loss component (\ref{['loss_analysis']}).
• Figure 5: Maximum mean discrepancy (MMD) comparison between ProULearn and MHPL during training on the Office-31 dataset (A$\rightarrow$D).