Abstract
Prototypical Networks (ProtoNet) are a foundational method for few-shot classification but suffer significant performance degradation in cross-domain scenarios. This failure is often attributed to "simplicity bias," where models latch onto easy, non-transferable shortcuts like background color or texture. We propose the Local-global Distillation Prototypical Network (LDP-net) to address this. LDP-net utilizes a dual-branch architecture with local-global distillation and cross-episode momentum updates to force the extraction of robust, invariant semantic features. Our method achieves state-of-the-art results across eight diverse target domains, demonstrating superior generalization compared to standard ProtoNet variants.
Introduction
Few-Shot Classification (FSC) aims to recognize novel categories from extremely limited data. While Prototypical Networks are effective within a single domain, they struggle with Cross-Domain Few-Shot Learning (CD-FSL), where training and testing data come from different distributions (e.g., natural images vs. satellite imagery). The primary hurdle is simplicity bias: deep networks tend to exploit simple shortcuts present in the training domain that do not generalize. To succeed, models must capture the core semantic "essence" of objects rather than superficial background cues.
Methodology: The LDP-net Framework
LDP-net moves beyond the single-stream architecture of traditional ProtoNets by introducing a dual-branch system designed to break simplicity bias through knowledge distillation.
Dual-Branch Architecture
The framework consists of a Global Branch, which processes full query images, and a Local Branch, which processes random local crops. By comparing predictions between these branches, the model is forced to find semantic meaning even in partial views of an object.
Knowledge Distillation (KD)
The model enforces consistency between the two branches through two types of distillation:
- Self-image Distillation: Ensures the local crop's prediction matches the global image's prediction, encouraging the network to identify semantic features within small regions.
- Cross-image Distillation: Aligns a local crop's prediction with the global prediction of a different image from the same class. This regularizer helps the model ignore intra-class variations like pose or lighting.
Momentum Updates
To stabilize learning across diverse meta-learning episodes, the local branch is updated as an Exponential Moving Average (EMA) of the global branch. This "cross-episode distillation" allows the model to accumulate long-term knowledge, preventing overfitting to the quirks of individual tasks.
Experimental Results
LDP-net was trained on mini-ImageNet and evaluated across eight diverse target domains, ranging from fine-grained natural images (CUB, Cars) to extreme domain shifts (ChestX, EuroSAT).
Quantitative Performance
LDP-net consistently outperformed baselines. On the CUB dataset (1-shot), it achieved 49.82% accuracy, a nearly 10% improvement over ProtoNet++ (40.34%). Significant gains were also observed in satellite (EuroSAT) and agricultural (CropDisease) imagery compared to previous methods like GNN+AFA and RDC.
Qualitative Analysis
Class Activation Maps (CAM) revealed that LDP-net focuses on the entire object rather than biased corners. Furthermore, visualization of the loss landscape confirms that LDP-net produces a smoother, wider landscape, which is strongly correlated with superior generalization capability.
Conclusion
LDP-net effectively mitigates simplicity bias in cross-domain few-shot learning by leveraging local-global distillation and momentum-based updates. While performance on extreme shifts like medical imaging remains a challenge, the robust semantic features learned by LDP-net represent a significant step forward for generalized few-shot classification. Future work will explore increasing training data diversity and more efficient target-domain adaptation strategies.
Summary updated