Abstract
Unsupervised Domain Adaptation (UDA) aims to bridge the gap between a labeled source domain and an unlabeled target domain. A major challenge in UDA is the accumulation of errors during iterative pseudo-labeling. This paper proposes a Selective Pseudo-Labeling (SPL) framework that leverages both source-centric Nearest Class Prototypes and target-centric Structured Prediction. By employing a progressive, class-wise selection strategy, the model avoids bias and error propagation. Experimental results on four major benchmarks—Office-Caltech, Office31, ImageCLEF-DA, and Office-Home—demonstrate that SPL achieves state-of-the-art performance.
Introduction
In machine learning, models often face "domain shift" when trained on a source dataset but deployed on a target dataset with different characteristics (e.g., studio photos vs. smartphone snapshots). Unsupervised Domain Adaptation (UDA) seeks to mitigate this performance drop without requiring target labels. A common technique is pseudo-labeling, where the model's own predictions on target data are used as training labels. However, this often leads to a "vicious cycle" where early errors are reinforced, causing catastrophic error accumulation. This paper addresses this challenge by introducing a robust framework for selective and structured pseudo-labeling.
Methodology
The proposed Selective Pseudo-Labeling (SPL) method is built upon three core components designed to ensure high-quality label assignment and stable adaptation.
1. Domain Alignment via SLPP
The framework utilizes Supervised Locality Preserving Projection (SLPP) to learn a shared subspace. In this space, samples from the same class are projected close together, regardless of whether they originate from the source or target domain.
2. Hybrid Pseudo-Labeling
To improve labeling accuracy, the authors combine two complementary perspectives:
- Nearest Class Prototype (NCP): A source-centric view that labels target samples based on their proximity to the mean features of source classes.
- Structured Prediction (SP): A target-centric view that clusters target data (using K-means) and matches these clusters to source classes via global optimization. This captures the intrinsic structure of the target domain.
3. Progressive Class-wise Selection
Rather than trusting all pseudo-labels at once, SPL adopts an "easy-to-hard" curriculum. It selects only the most confident samples in each iteration. Crucially, this selection is performed class-wise, ensuring the model maintains a balanced representation and does not become biased toward easily distinguishable classes.
Key Insights
- Structural Awareness: Target data possesses internal clusters. By using Structured Prediction, the model identifies these clusters globally rather than treating points in isolation, leading to more reliable labels.
- Complementarity: NCP excels for target samples similar to the source, while SP is more effective for samples that are far from the source but reside within well-defined target clusters.
- Balanced Learning: Class-wise selection prevents the model from ignoring difficult classes, which is a common failure mode in simple confidence-based selection strategies.
Experimental Results
The SPL method was evaluated on four standard benchmarks: Office-Caltech, Office31, ImageCLEF-DA, and Office-Home.
- Office-Caltech: Achieved 93.0% average accuracy.
- Office31: Reached 89.6%, outperforming several deep learning-based methods.
- Office-Home: On this complex 65-category dataset, SPL achieved 71.0%, a significant improvement over previous state-of-the-art methods like TADA and SymNets (67.6%).
Ablation studies confirmed that both the Structured Prediction component and the Selective strategy are vital for performance. As shown in the sensitivity analysis, the model remains relatively stable across different hyperparameter settings, such as the dimensionality of the projection spaces and the number of iterations.
Conclusion
SPL provides a robust framework for UDA by combining structured prediction with a selective learning strategy. While currently implemented as a subspace learning technique on top of deep features, future work could involve integrating this logic directly into end-to-end deep neural network training to allow for more fluid feature adaptation.
Summary updated