Techniques & Architectures

Few-Shot Learning

Few-shot learning is a machine learning approach where models are designed to learn from a very limited number of training examples, typically between 1-5 samples per class, whilst maintaining good performance.

Few-shot learning attempts to mimic the human ability to learn new concepts quickly with minimal examples. The approach leverages prior knowledge and learning strategies to make accurate predictions despite having access to only a small number of labelled training samples for new tasks or categories.

This learning paradigm is particularly valuable in scenarios where collecting large amounts of training data is expensive, time-consuming, or impossible. Few-shot learning models often employ meta-learning techniques, where the model “learns how to learn” from a variety of related tasks, enabling it to adapt quickly to new situations with minimal data.

Examples

  • Medical diagnosis with rare conditions
  • Face recognition with limited photos
  • Species classification from few specimens
  • Product detection with minimal samples
  • Handwriting recognition from few examples
Related Terms
×