Topic 12 of 20 · Part III : Personalisation and Responsible AI

Active learning

Active learning chooses which production examples deserve human review instead of labelling data at random.

The control pattern

Data asset

Selected production examples for human labelling

Knobs

Uncertainty threshold, Disagreement, Sampling method, Business impact, Novelty, Review priority

Outcome

Faster improvement with fewer labelled examples

Measurement

See how to measure this below

What it is

Active learning chooses which production examples deserve human review instead of labelling data at random. The data asset is the stream of uncertain, novel or high-impact cases; the knobs determine how examples are selected.

Common signals include low confidence, disagreement between models, failure of a validator, unusual input structure, user correction and high business consequence. The method is most effective when the labelling budget is limited and the current model already handles ordinary cases well.

Reviewed examples should feed several destinations: a corrected production record, the evaluation suite, retrieval rules, prompt examples, and eventually fine-tuning or distillation data.

The knobs in detail

Each row is one adjustable property of the data asset, and what moving it tends to do.

KnobWhat you adjustLikely effect
Uncertainty thresholdHow unsure a case must be to queueFocuses review where the model is weakest
DisagreementConflict between models or validatorsSurfaces silent errors
Sampling methodHow the queue is composedDetermines what the team ever sees
Business impactFinancial materiality of the casePrioritises consequential errors
NoveltyHow unlike the training set the input isCatches emerging document types
Review priorityOrder of the queueShortens time to fix

Applied: Stocks Assistant

For Stocks Assistant, a review queue might prioritise conflicting EPS values across sources, a newly introduced filing table, ambiguous ticker resolution, an earnings summary rejected by a user, or a high-confidence recommendation contradicted by a verifier. Examples can be scored on uncertainty, novelty, expected learning value and financial materiality.

How to measure it

Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.

  • Label turnaround time
  • Reviewer agreement
  • Error reduction per labelled example
  • Recurrence rate of the targeted failure

Common mistakes

Queueing only known weaknesses. Reserve space for random cases, or you will miss silent failures where the model is confidently wrong.

Retraining on unverified feedback rather than reviewed corrections.