Introduction
Few-shot learning (FSL) aims to develop models capable of recognizing novel categories from a limited number of labeled examples. While significant progress has been made in intra-domain settings, performance often degrades sharply under domain shift—where the source training data (e.g., general objects) differs significantly from the target test data (e.g., medical imaging or satellite photos).
Existing approaches typically fall into two categories: meta-learning, which learns generalizable learning strategies but often overfits to source-domain features, and transfer learning, which adapts well to new domains but fails to exploit the relational structure within the provided support set. DAMSL (Domain Agnostic Meta Score-based Learning) bridges this gap by combining the adaptability of transfer learning with the relational modeling of metric-based meta-learning.
Methodology: Score-Based Metric Learning
The core innovation of DAMSL is the shift from high-dimensional image features to Score-based Metric Learning. By using classification scores (logits) as coordinates in a new metric space, the model abstracts away domain-specific visual noise, creating a more domain-agnostic representation.
The DAMSL framework operates in three distinct stages:
- Fine-tuning: A ResNet-10 backbone is fine-tuned on the target domain's support set to generate initial class scores.
- Domain Agnostic Coordinates: These scores serve as inputs to the metric learner. The authors argue that the relative relationships between scores (e.g., how much an image resembles various classes) are more consistent across domains than raw pixel features.
- Graph Neural Network (GNN) Refinement: A GNN treats each image as a node and refines the classification by analyzing the entire distribution of the support set. This allows the model to "correct" initial biases from the fine-tuned encoder by leveraging the structural relationships between all available samples.
Experimental Results
The authors evaluated DAMSL on the BSCD-FSL (Broader Study of Cross-Domain Few-Shot Learning) benchmark, which includes diverse datasets such as CropDisease (plant pathology), EuroSAT (satellite imagery), ISIC (skin lesions), and ChestX (chest X-rays).
Key Findings
- State-of-the-Art Performance: DAMSL v2 achieved an average accuracy of 74.99% across the benchmark, representing a 6.86% improvement over the TransFT baseline and a 15.21% lead over ProtoNet.
- Robustness to Shot Count: The model demonstrated consistent superiority across 5-shot, 20-shot, and 50-shot scenarios, showing that the GNN-based refinement scales effectively with more data.
- Ensemble Benefits: DAMSL v2, which utilizes an ensemble of backbones trained with different optimization strategies (Adam and SGD), outperformed the single-backbone DAMSL v1, suggesting that diverse score coordinates provide a richer foundation for the GNN.
Conclusion and Future Work
DAMSL demonstrates that moving the meta-learning process from the feature level to the score level effectively mitigates the impact of domain shift. By leveraging the relational power of GNNs on top of fine-tuned logits, the framework achieves remarkable consistency across highly varied datasets.
Despite these gains, extreme domain shifts—exemplified by the ChestX dataset—remain a challenge for the field. Future research may explore integrating DAMSL with even larger pre-trained models or applying its score-based refinement as a modular "self-correction" layer for broader robust classification tasks.
Summary updated