Introduction
Cross-Domain Few-Shot Learning (CD-FSL) represents a significant hurdle in computer vision. While standard Few-Shot Learning (FSL) assumes that training (source) and testing (target) data share a distribution, real-world applications often involve substantial domain shifts—such as moving from natural photographs to sketches or thermal imagery. This paper introduces the Discriminativeness-Preserved Domain Adaptive Prototypical Network (DPDAPN), a unified framework designed to bridge this gap by balancing domain alignment with class discriminativeness.
The Conflict: Alignment vs. Discrimination
The primary challenge in CD-FSL is the inherent conflict between Domain Adaptation (DA) and FSL objectives. Standard DA techniques aim to align global distributions so that the model cannot distinguish between domains. However, aggressive alignment can lead to "over-alignment," where the fine-grained features necessary to distinguish between classes are lost. In FSL, where the model must generalize to unseen classes with very few examples, maintaining high "discriminativeness" is critical. DPDAPN addresses this by explicitly preserving class separation before performing domain alignment.
The DPDAPN Framework
The DPDAPN architecture integrates three core components to resolve the alignment-discrimination trade-off:
- Prototypical Network Base: The model employs a metric-learning approach, calculating "prototypes" (mean vectors) for each class and classifying new instances based on their proximity to these prototypes.
- Domain Adversarial Adaptation: To minimize the domain gap, the framework utilizes Margin Disparity Discrepancy (MDD). This involves a minimax game where the network attempts to maximize discrepancy between source and target episodes while the feature embedding works to minimize it.
- Domain Discriminative Loss: This is the framework's "secret sauce." By applying a specific loss to the feature extractor's output before the final embedding layer, the model is forced to retain features useful for class separation, even as it learns to ignore domain-specific noise.
- Adaptive Reweighting Module (ARW): To manage the complexity of balancing four distinct losses (two for FSL and two for DA), the authors implemented an automated weighting scheme. This module uses task-dependent uncertainty to dynamically prioritize objectives during training, ensuring stability without manual hyper-parameter tuning.
Experimental Results
The authors evaluated DPDAPN across several benchmarks, including miniImageNet, tieredImageNet, and the challenging DomainNet.
- miniImageNet: In 5-way 1-shot tasks with style-transferred target domains, DPDAPN achieved 30.23% accuracy, outperforming standard Prototypical Networks (24.68%) and naive DA+FSL combinations (27.12%).
- DomainNet: On the "Real to Sketch" shift, DPDAPN reached 66.86% in 5-way 5-shot tasks, a significant 7% improvement over the next best baseline.
Ablation studies confirmed that the Domain Discriminative loss is the primary driver in preventing performance degradation during the alignment process, effectively buffering class-specific features against the effects of global distribution matching.
Conclusion
DPDAPN demonstrates that global domain alignment is most effective for few-shot tasks when paired with explicit safeguards for class discriminativeness. By treating DA and FSL as a joint optimization problem and utilizing uncertainty-based reweighting, the framework provides a robust solution for transferring knowledge across disparate visual domains. Future work may explore more sophisticated definitions of discriminativeness to handle even more extreme domain shifts, such as those found in medical or satellite imagery.
Summary updated