Guidance

Machine learning with limited data

A guide to machine learning techniques for limited data problems, including approaches for small amounts of data and for large amounts of unlabelled data.

Documents

Machine Learning with Limited Data - original version

Request an accessible format.
If you use assistive technology (such as a screen reader) and need a version of this document in a more accessible format, please email centralenquiries@dstl.gov.uk. Please tell us what format you need. It will help us if you say what assistive technology you use.

Details

Machine learning is a branch of artificial intelligence (AI) where computers algorithms examine datasets, find common patterns, and learn and improve without being explicitly programmed. It offers Defence significant benefits, such as increasing the amount of data that can be analysed and reducing the load on human analysts. Sometimes state-of-the-art machine learning models cannot be applied due to a lack of data or the expense and time required to label enough examples, so other approaches must be taken.

This handbook informs and guides machine learning practitioners on the machine learning approaches currently possible for solving limited data problems, including:

  • how much data is required
  • ways to deal with small amounts of data, including zero-shot learning and meta-learning
  • ways to deal with large amounts of mostly unlabelled data
  • model refinement
  • technology readiness levels

This guide captures some of the knowledge gained from Dstl research into low-shot learning carried out by the Future of AI for Defence project, part of the Dstl Autonomy programme.

Updates to this page

Published 7 December 2020

Sign up for emails or print this page