Abstract
Few-Shot Learning (FSL) often suffers from data scarcity in the source domain, particularly in fine-grained classification tasks. While web-crawled data can supplement training sets, it introduces significant noise. This paper proposes AdarGCN, a Graph Convolutional Network (GCN) architecture utilizing adaptive, multi-level aggregation to filter noise and prevent over-smoothing. By dynamically weighting information propagation across different neighborhood scales, AdarGCN achieves state-of-the-art performance on benchmarks like mini-ImageNet and CUB.
Introduction
Standard Few-Shot Learning assumes abundant "source" classes for pre-training. However, in specialized domains like fine-grained bird recognition, even source data is limited—a problem termed Scarce-Source Few-Shot Learning (SSFSL). Web crawling offers a solution but brings "noisy" data (e.g., irrelevant images). Traditional GCNs propagate this noise indiscriminately and suffer from over-smoothing in deeper layers, where node features become indistinguishable. AdarGCN addresses these issues by automatically learning how far information should propagate between nodes.
Methodology
The AdarGCN framework operates in a two-stage process designed to leverage noisy web data effectively.
Two-Stage Framework
- Stage 1 (Label Denoising - LDN): AdarGCN is first used to classify web-crawled images as "relevant" or "irrelevant" to clean the dataset.
- Stage 2 (Few-Shot Learning): The cleaned images are merged with the original data, and a second AdarGCN is trained to perform the final classification task.
Adaptive Aggregation Mechanism
The core innovation is the Adaptive Aggregation module within the GCN layers. Unlike fixed neighborhood updates, AdarGCN employs parallel branches to aggregate information at varying scales:
- Branch c: Self-features (0-hop).
- Branch b: Immediate neighbors (1-hop).
- Branch d: Extended neighborhood (2-hop).
A learnable gating mechanism assigns weights to these branches for every node. This allows the model to "quarantine" noisy nodes by prioritizing their own features over neighborhood information, effectively acting as a spatial filter.
Key Results
AdarGCN was evaluated on mini-ImageNet and the CUB (Caltech-UCSD Birds-200-2011) fine-grained dataset.
- Label Denoising: On mini-ImageNet with only 10 clean samples, AdarGCN achieved 63.37% accuracy, significantly outperforming standard GCNs (61.22%) and Label Propagation (60.21%).
- Conventional FSL: Even without web-crawling, AdarGCN sets a new state-of-the-art for GCN-based few-shot learning. On mini-ImageNet (5-way, 5-shot), it reached 71.48% accuracy, beating previous GCN methods like EGCN (66.85%).
- Fine-Grained Success: On the CUB dataset, the model achieved 78.04% accuracy in the 5-shot setting.
Discussion and Conclusion
AdarGCN effectively addresses the SSFSL problem by treating aggregation depth as a learnable parameter. Its architecture mitigates over-smoothing, allowing for deeper GCNs (up to 7 layers) that maintain high performance. While the multi-head structure adds computational complexity, the gains in robustness to web noise make it a powerful tool for data-scarce scenarios. Future work may explore applying this adaptive logic to other graph-based tasks like recommendation systems or social network analysis.
Summary updated