Skip to content

Prototypical Networks for Few-shot Learning

Authors: Jake Snell, Kevin Swersky, Richard Zemel Year: 2017 Venue: NeurIPS 2017 Citations: 10,000+

Summary

Introduces Prototypical Networks for few-shot learning. Classifies examples by computing distances to prototypes (class means in embedding space). Simple, intuitive approach to learning from few examples via metric learning.

Key Concepts

  • Metric Learning: Learn embedding space where distances reflect similarity
  • Prototypes: Class means in learned embedding space
  • Few-Shot Learning: Classify with few examples per class
  • Distance-Based Classification: Assign to nearest prototype
  • Episodic Training: Train on tasks sampling from class distribution
  • Transductive Setting: Use query set statistics during testing

Impact

  • 10,000+ citations
  • Simple and intuitive few-shot learning approach
  • Strong baseline for few-shot classification
  • Influenced many metric learning approaches
  • Practical and well-understood method
  • Applications in zero-shot and few-shot scenarios
  • Foundation for understanding similarity in learning

Key Results

  • Few-shot classification with simple distance metric
  • Competitive with more complex approaches
  • Easy to understand and implement
  • Good empirical performance
  • Matching Networks (Vinyals et al., 2016)
  • Relation Networks (Sung et al., 2018)
  • Model-Agnostic Meta-Learning (Finn et al., 2017)
  • Siamese Networks (Koch et al., 2015)