Think Build Implement Repeat
London, UK +44 7367 067226
WhatsApp FOLLOW f in X
  1. Home
  2. Blog
  3. Active Learning: Labelling Fewer Examples, Better
AI & Machine Learning

Active Learning: Labelling Fewer Examples, Better

Not every labelled example is equally useful. How to let the model choose what to label next, and cut labelling effort substantially.

Updated 2 min readBy SpiderHunts Technologies

Free estimateNo obligation

Get a free estimate

Tell us what you need. A senior engineer reads every enquiry.

Takes under a minute. We never share your details.

  • Free consultation
  • No commitment
  • NDA on request

Prefer to talk? Book a free 30-minute call →

Quick answer — TL;DR

Random sampling wastes labelling effort on examples the model already handles. Active learning selects the cases it is least certain about, which typically reaches a given accuracy with substantially fewer labels - though it needs guarding against a sample that drifts unrepresentative.

Most labels teach nothing

If you label a thousand random examples, a large share will be cases the model already classifies confidently and correctly. They cost the same to label as the difficult ones and contribute far less.

Active learning inverts the order: train on a small set, ask the model which unlabelled examples it is least sure about, label those, retrain, repeat. The labelling effort concentrates on the boundary where the decisions are actually hard.

How a round works in practice

  1. Label a modest random seed set - this has to be random, to establish an unbiased picture.
  2. Train an initial model, however mediocre.
  3. Score the unlabelled pool and select where confidence is lowest.
  4. Label that batch, add it, retrain.
  5. Repeat until accuracy on a held-out set stops improving materially.

Batch sizes matter practically. Labelling one example at a time is statistically neat and operationally absurd; a batch of fifty to a few hundred fits how people actually work.

The traps

  • The labelled set stops being representative. It is deliberately concentrated on hard cases, so accuracy measured on it is misleading. Always keep a separate random test set.
  • Rare categories get missed. If the model has never seen a category it cannot be uncertain about it. Mix in some random or targeted sampling.
  • Uncertainty is misread. An example may be uncertain because it is genuinely ambiguous rather than informative - those add noise, not signal.
  • Labelling gets harder. Every batch is drawn from the difficult end, which is tiring and slower per item. Set expectations with the people doing it.

When it is worth the machinery

Active learning adds a loop between labelling and training, which is operational complexity. It pays when labelling is expensive - expert time, medical or legal review - or when the unlabelled pool is enormous relative to what you can afford to label.

SituationWorth it?
Expert labelling at high costUsually yes
Millions of unlabelled recordsUsually yes
A few thousand records, cheap to labelProbably not - just label them
Categories still being definedNo - settle the taxonomy first

A simpler variant that often suffices

Full active learning is not always necessary. A single targeted round - train on what you have, find where the model is weakest, label more of that - captures much of the benefit at a fraction of the complexity.

This is particularly effective where one category is underperforming. Rather than labelling more of everything, label more of that, which is an obvious idea that projects routinely fail to act on.

The examples worth labelling are the ones the model gets wrong, not the ones it finds easy.

FAQ

Frequently asked questions

The questions readers ask us after this guide.

Still have a question?

Ask us directly — a senior engineer will get back to you.

Ask about your project

How much labelling effort does this save?

It varies with the problem and cannot be promised in advance. The saving is largest where the decision boundary is a small part of a large pool.

Do we still need a random test set?

Yes, and it must stay random. The actively selected set is deliberately unrepresentative and cannot measure real-world accuracy.

Can this be combined with model-assisted labelling?

Yes, and they work well together - the model proposes a label and selects which examples to propose on, with humans correcting.

Does it work for regression as well as classification?

The principle transfers, using prediction interval width rather than class confidence, though it is less commonly applied.

Keep reading

More on AI & Machine Learning

Start here

Want machine learning project details from us?

Tell us what you are trying to predict and roughly what data you hold. We will come back with an honest view on whether machine learning is the right tool, what the work would involve and a realistic cost range. If a spreadsheet would do the job, we will say so.

  1. You tell us what you needTwo minutes on the form, or a message on WhatsApp.
  2. A senior engineer reviews itAnd comes back with questions, a realistic range and an honest view on fit.
  3. Free 30-minute scoping callWe talk through scope, options and a realistic estimate — with no obligation.
Free estimateNo obligation

Talk to someone who builds this

Send a short brief and we will come back with an honest view and a realistic range.

Takes under a minute. We never share your details.

  • Free consultation
  • No commitment
  • NDA on request

Prefer to talk? Book a free 30-minute call →