# Prototypical Networks for Few-shot Learning

**Authors:** Jake Snell, Kevin Swersky, Richard S. Zemel  
**arXiv:** [1703.05175](https://arxiv.org/abs/1703.05175)  
**Date:** 2017-03-15  
**Summary page:** https://www.summarizepaper.com/en/paper/1703.05175v1/

## Key points

- Authors: Jake Snell, Kevin Swersky, Richard S. Zemel
- Introduce prototypical networks for few-shot learning
- Operate by learning a metric space and calculating Euclidean distances to prototype representations of each class
- Simpler inductive bias compared to recent approaches for few-shot learning
- Achieve state-of-the-art results in scenarios with limited data
- Extend application to zero-shot learning, achieving state-of-the-art results on the CU-Birds dataset

## Summary

, , , , 

In their paper titled "Prototypical Networks for Few-shot Learning," authors Jake Snell, Kevin Swersky, and Richard S. Zemel introduce prototypical networks as a solution to the problem of few-shot classification. This task involves training a classifier to generalize to new classes not present in the training data, with only a small number of examples available for each new class. Prototypical networks operate by learning a metric space where classification is achieved by calculating Euclidean distances to prototype representations of each class. The key innovation of prototypical networks lies in their simpler inductive bias compared to recent approaches for few-shot learning. This simplicity proves advantageous in scenarios with limited data, enabling prototypical networks to achieve state-of-the-art results. The authors conduct an analysis that demonstrates how straightforward design choices can lead to significant improvements over more complex architectures and meta-learning techniques commonly used in the field. Furthermore, the authors extend the application of prototypical networks to zero-shot learning, where models are required to classify objects from classes they have never seen before. Remarkably, their approach achieves state-of-the-art results in zero-shot learning on the CU-Birds dataset. Overall, this research showcases the effectiveness of prototypical networks in addressing challenges related to few-shot and zero-shot learning tasks. By emphasizing simplicity and leveraging prototype representations, these networks offer a promising avenue for advancing machine learning capabilities in scenarios with limited labeled data.