Abstract
Few-shot classification (FSC) aims to recognize novel categories from extremely limited data. While self-supervised learning (SSL) is often used to enhance feature representations, it frequently suffers from representation bias in data-scarce regimes. This paper introduces Conditional Self-Supervised Learning (CSS), a three-stage framework that leverages supervised labels to guide self-supervised feature discovery. By making SSL conditional on supervised priors, CSS prevents the model from learning irrelevant "shortcuts" and achieves state-of-the-art performance on standard FSC benchmarks without requiring additional unlabeled data.
Introduction
The core challenge of few-shot classification is learning robust, generalizable features from only a few examples. To bridge the gap between human-like learning and deep learning models, researchers have integrated self-supervised learning (SSL) as an auxiliary task. However, standard SSL techniques are typically data-hungry. In the low-data regime of FSC, SSL can "go rogue," focusing on low-level patterns like color histograms or edge continuity rather than semantic information. This representation bias can degrade performance on the primary classification task.
To address this, the authors propose Conditional Self-Supervised Learning (CSS). The key idea is to use supervised information as a "prior" to constrain the self-supervised branch, ensuring it explores feature variations that are useful for the downstream task.
Methodology
The CSS framework consists of a three-stage training pipeline designed to harmonize supervised and self-supervised objectives.
1. Supervised Pre-training
In the first stage, a standard feature extractor $f_\theta$ is trained using all available labeled data from the base classes. This establishes a baseline supervised representation.
2. Conditional Self-Supervised Training
This stage introduces a second feature extractor $g_\xi$, trained via a self-supervised method (e.g., SimSiam). Unlike vanilla SSL, a Conditional Module forces $g_\xi$ to align with the supervised features from $f_\theta$. By minimizing the distance between these representations ($L_{cond}$), the supervised model acts as a teacher, guiding the self-supervised branch to ignore irrelevant shortcuts and focus on meaningful semantic structures.
3. Meta-training with Fusion Distillation
Finally, the knowledge from both the supervised ($f_\theta$) and self-supervised ($g_\xi$) branches is fused into a final embedding network $h_\phi$. A Fusion Distillation (FD) method, utilizing a Graph Convolutional Network (GCN), aggregates information across samples to enrich the final feature representation used for meta-learning.
Experimental Results
The CSS model was evaluated on three standard benchmarks: CIFAR-FS, CUB-200 (fine-grained birds), and mini-ImageNet.
Key Performance Metrics
CSS consistently achieved state-of-the-art results in both 1-shot and 5-shot scenarios:
- mini-ImageNet: Reached 68.08% accuracy in 5-shot tasks, significantly higher than traditional Prototypical Networks (64.69%).
- CUB-200: Achieved 81.84% in 5-shot tasks, outperforming other SSL-based FSC methods by 5-10%.
- CIFAR-FS: Reached 74.59% in 5-shot tasks.
Insights
Ablation studies confirmed that "vanilla" self-supervision often performs worse than simple supervised learning in few-shot settings. The Conditional Module was identified as the primary driver of performance gains, proving that guided exploration is essential when data is limited. The best results were obtained by integrating both paradigms: the supervised branch captures inter-class differences, while the self-supervised branch mines intra-sample structures.
Conclusion and Future Work
CSS demonstrates that self-supervision, when properly conditioned on supervised priors, can significantly boost few-shot classification performance. By preventing representation bias, the framework ensures that learned features are both semantically rich and task-relevant.
Future research directions include addressing the computational overhead of the three-stage process and the GCN fusion. Additionally, exploring automated hyperparameter tuning for the loss balancing and scaling the framework to larger backbones like Vision Transformers (ViTs) could further enhance its utility in diverse computer vision applications.
Summary updated