Introduction
Standard Few-Shot Learning (FSL) typically assumes that while the target classes have very few examples (e.g., 1 or 5 images), a massive, clean source dataset is available to pre-train the model. However, in specialized domains like fine-grained bird recognition, even source data can be scarce. This paper introduces a more realistic and challenging setting called Few-Shot Few-Shot Learning (FSFSL), where data is limited for both source and target classes.
A common strategy to mitigate data scarcity is to crawl images from the web (e.g., Google Image Search). However, web data is inherently noisy, often containing irrelevant images that can degrade model performance. To address this, the authors propose AdarGCN (Adaptive Aggregation Graph Convolutional Network), a novel architecture designed to filter noise and learn robust representations through dynamic information flow.
The AdarGCN Framework
The proposed solution operates in a two-stage pipeline powered by the AdarGCN architecture:
- Stage 1: Label Denoising (LDN): Images are crawled from the web and processed by AdarGCN acting as a binary classifier. It distinguishes between "clean" images belonging to the class and "noisy" irrelevant images.
- Stage 2: Few-Shot Learning: The original clean data is combined with the denoised web data to train the final FSL model using the same AdarGCN architecture.
Adaptive Aggregation Mechanism
Traditional GCNs propagate information uniformly across a graph, which can lead to "over-smoothing" or the propagation of noise from outlier nodes. AdarGCN introduces a multi-head multi-level aggregation module that allows each node to have a dynamic "aggregation depth."
As shown in the architecture details, the model processes node features through three distinct branches:
- Branch C: No aggregation (the node feature remains unchanged).
- Branch B: One iteration of neighbor aggregation.
- Branch D: Two iterations of aggregation, looking further into the graph structure.
A learnable gate dynamically weights these branches. For instance, if a node is identified as a noisy outlier, the model can prioritize Branch C to ignore its neighbors. Conversely, for high-quality samples in dense clusters, it can aggregate more broadly using Branch D.
Experimental Results
AdarGCN was evaluated on standard benchmarks including mini-ImageNet and CUB (Caltech-UCSD Birds-200-2011). The results demonstrate significant improvements over state-of-the-art methods:
- Label Denoising: On mini-ImageNet, AdarGCN achieved a denoising accuracy of 63.37% with only 10 clean samples, outperforming standard GCN and ResGCN baselines.
- FSFSL Performance: In the challenging FSFSL setting on the CUB dataset, the model reached 80.88% accuracy, a substantial jump over established baselines like MAML and ProtoNet.
- Standard FSL: Even when source data is plentiful, AdarGCN achieved 71.48% on mini-ImageNet, surpassing previous GCN-based methods by approximately 5%.
Conclusion
AdarGCN effectively addresses the challenges of data scarcity and web-data noise in few-shot learning. By employing adaptive aggregation, the model prevents "bad" information from noisy nodes from polluting the graph while allowing for deeper, more robust GCN architectures. Future work may explore scaling these benefits with more complex feature backbones and addressing open-set noise where crawled images belong to entirely unseen classes.
Summary updated