Rethinking Class Relations: Absolute-relative Few-shot Learning

Hongguang Zhang, Philip H. S. Torr, Hongdong Li, Songlei Jian, Piotr Koniusz

TL;DR

Current few-shot learning models rely on simplistic binary labels (same/different) to learn image relations, ignoring the rich semantic nuances between objects. This paper introduces Absolute-relative Learning (ArL), a paradigm that forces networks to simultaneously learn what an object is (absolute) and how it semantically relates to others (relative). By using attributes and self-supervised augmentations as auxiliary training signals, ArL significantly boosts both supervised and unsupervised few-shot classification performance.

The Problem with Binary Relations

Few-shot learning (FSL) aims to recognize new categories from just one or a handful of examples. A popular approach is "relation learning," where a network learns to compare a query image against support images. However, these networks are typically trained using rigid binary labels: 1 if images are from the same class, and 0 if they aren't.

Conceptual comparison of Absolute, Relative, and Absolute-relative Learning paradigms.

The authors identify two major flaws with this binary approach:

  1. Lack of Decision Smoothness: Binary labels treat all "different" pairs equally. However, a tiger and a house cat are visually more similar than a tiger and a car. Failing to capture these nuances leads to brittle decision boundaries.
  2. Ignoring the Object Concept: Because the network only outputs a similarity score, it never explicitly learns what the objects are, discarding valuable class-concept knowledge.

The Absolute-relative Learning (ArL) Framework

To address these issues, the authors propose Absolute-relative Learning (ArL). This multi-task learning setup, applied during training, consists of two main modules attached to a standard feature encoder:

  • Absolute Learning (AL): An auxiliary branch forces the network to predict explicit object information from a single image, such as class labels or semantic attributes (e.g., "furry," "has legs"). This ensures the encoder learns concrete object concepts.
  • Relative Learning (RL): Instead of binary similarity, this module predicts "soft" semantic similarities based on shared attributes or word2vec embeddings.
  • The Feedback Loop: Crucially, predictions from the Absolute Learner are fed back into the Relative Learner to refine the final similarity prediction, creating a more robust feature space.
Detailed architecture of the ArL framework, showing the interaction between absolute and relative learning branches and the feedback mechanisms.

Unsupervised Extension

ArL can also operate without labels by using image augmentations (rotations, flips, color jitter). In this mode, the AL module predicts which augmentation was applied, while the RL module predicts the similarity between different augmented views of the same image, effectively integrating contrastive learning principles.

Key Insights and Results

ArL acts as a plug-and-play module that improves various baseline architectures, including Prototypical Nets and RelationNet. Key findings include:

  • Attributes > Word Embeddings: For fine-grained tasks (like distinguishing bird species), visual attributes provide a much cleaner training signal than general word2vec embeddings.
  • Supervised Performance: On the miniImageNet benchmark, adding ArL to the Second-order Similarity Network (SoSN) improved 1-shot accuracy by 3.6% (reaching 61.36%) and 5-shot by 4.1% (reaching 78.95%).
  • Unsupervised Gains: The unsupervised variant achieved massive gains, particularly on fine-grained datasets like Flower102, where it boosted 1-shot accuracy by 7.9% over the baseline. Remarkably, unsupervised ArL often outperformed older supervised methods.

Conclusion

By moving beyond binary similarity and forcing models to learn both object concepts and semantic relations, ArL creates smoother and more robust representations for few-shot learning. Because semantic information is used only as a training target, the model requires no extra annotations during inference, making it a practical and powerful enhancement for existing FSL architectures.

Summary updated

BibTeX
@inproceedings{zhang2021rethinking,
  author = {Hongguang Zhang and Philip H. S. Torr and Hongdong Li and Songlei Jian and Piotr Koniusz},
  title = {{Rethinking Class Relations: Absolute-relative Few-shot Learning}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2021},
  doi = {10.1109/cvpr46437.2021.00931},
}
Figure