AI Glossary / Governance, ethics and safety / Advanced

Adaptive learning

Definition

Adaptive learning, in the machine-learning sense used here, updates a predictive model during operation so that it can respond to changing patterns in incoming data. It is especially relevant when relationships learned from earlier observations become less useful. The phrase also appears in education for personalized instruction; that application should not be confused with the broader model-updating mechanism.

What it means

A model trained once carries assumptions from its training period. If customer behavior, equipment conditions or document vocabulary changes, those assumptions may become stale. Adaptive learning provides a way to revise learned behavior as further evidence becomes available.
Adaptation is not synonymous with improvement. A system can learn from noisy, incomplete or misleading feedback and become worse. Nor does a changing prediction prove that learning occurred: a fixed model naturally produces different answers for different inputs. The relevant question is whether the system's learned state changes and how that change affects subsequent predictions.

How it works

A supervised adaptive system makes a prediction, records the relevant input, and later receives an observed outcome. It can then update model parameters using that outcome. Some methods update continuously; others use a change detector to trigger a replacement or adjustment. Recent observations may receive greater influence than older ones.
Incremental learning supplies the ability to learn in successive pieces, but that ability alone does not guarantee effective responses to drift. A learner that retains all old evidence equally may respond too slowly. Conversely, forgetting quickly can discard useful patterns that later return.
Evaluation must preserve the information available at prediction time. Predict first, score against the outcome when available, and only then use that outcome for learning. If labels arrive days later, a realistic replay should reproduce that delay. Training on tomorrow's answers before assessing today's predictions gives a misleading result.

Three practical examples

• A help-desk classifier learns from reviewed routing corrections as new product names appear. Its team checks whether improvements for the new vocabulary come at the expense of older request categories.
• A machine-duration predictor updates from completed jobs after a workshop changes materials. It keeps timestamps because a duration becomes known only when a job finishes, not when it begins.
• A recommendation system adjusts learned preferences from interaction feedback. A team separately examines whether changes reflect users' interests or merely the restricted selection of items the system previously showed them.

Think of it this way

Think of revising a working estimate as new measurements arrive. Keeping every old measurement equally influential makes the estimate steady but potentially slow to move. Favoring recent measurements makes it responsive but more sensitive to unusual events. The appropriate balance depends on what is changing and how reliable the measurements are.

Worked example

Consider a deliberately simple predictor of job duration. Before a completed job is observed, its prediction is 10 minutes. After receiving the actual duration, it updates using:
New estimate = old estimate + 0.25 × (observed duration − old estimate).
The next three jobs each take 14 minutes, with the result of each available before the following prediction. The first prediction is 10, its absolute error is 4, and the updated estimate becomes 11.
The second prediction is 11, its error is 3, and the estimate becomes 11.75. The third prediction is 11.75, its error is 2.25, and the estimate becomes 12.3125.
The mean absolute error across these three predictions is (4 + 3 + 2.25) / 3, approximately 3.08 minutes. A frozen predictor that always outputs 10 would have a mean absolute error of 4 minutes on this particular sequence.
The final 12.3125 estimate must not be substituted for the earlier predictions when calculating that historical score. This illustration shows recency-weighted adaptation, not a universal learning algorithm or proof that updating always wins. One exceptional long job could also pull the estimate upward when no lasting change had occurred.

Illustrative case study

Fictional Willow Repairs introduces an adaptive job-duration estimator after adding a new repair service. An initial retrospective test looks excellent because it lets the model learn each completed job's duration before scoring its prediction for that same job.
The analyst corrects the order and replays the record using actual completion times. Performance is less impressive, particularly when several jobs begin before any of their outcomes are available. The team compares the adaptive estimator with a frozen baseline and reports recent errors separately for established and new services.
It also saves the update history so that an unexpected prediction can be traced to the information available then. The exercise reveals a useful adaptation opportunity while preventing an evaluation mistake from becoming an unsupported business claim.

Common misconceptions and mistakes

Assuming all AI products learn from every interaction: many deployed models remain fixed between managed updates.
Confusing data change with proven performance deterioration: changed inputs warrant investigation, but their practical effect must be measured.
Equating fast updates with good updates: noisy feedback can make a responsive model unstable.
Scoring predictions after learning their answers: this contaminates the evaluation.
Treating an adaptation mechanism as a complete governance process: responsibilities, monitoring and recovery still need to be defined.

Business use

Ask what can change, which observations authorize an update, how quickly reliable outcomes arrive, and how performance is compared with a stable reference. Keep enough records to reconstruct important predictions. Decide in advance what evidence should lead to continued adaptation, a paused update process or restoration of an earlier model.

People also ask

Does adaptive mean online?

Online updating is a common mechanism, but terminology varies across applications.

Is an adaptive learning platform necessarily retraining its model continuously?

No; personalized content selection may use fixed rules or models.

Does a drift alarm prove the cause of change?

No. Investigation is still needed.

Check your understanding

When should the first job's outcome enter the example's score?

Answer: After its original prediction has been recorded.

What does the 0.25 setting control?

Answer: The influence of the newest observation on the estimate.

Why retain a frozen baseline?

Answer: It helps assess whether updating actually adds value under the observed conditions.

Related terms

• Online learning
• Concept drift
• Data drift
• Feedback loop
• Continual Learning
• incremental learning
• progressive validation
• delayed labels

Learn next

Replay a short sequence in time order. Record each prediction before updating, compare it with a frozen baseline, and repeat with delayed outcomes to see how information timing changes the result.

More in Governance, ethics and safety

• Accountability - Accountability in AI means that identifiable people and organizations can be called upon to explain and justify decisions within their roles, demonstrate...
• Activity Bias - Activity bias occurs when data or conclusions disproportionately reflect highly active users while less active or inactive users are poorly represented.
• Adaptive dynamic programming - Adaptive dynamic programming is a model-based reinforcement-learning approach that learns how an environment behaves from experience and uses dynamic...
• Adversarial Autoencoders - Adversarial autoencoders are autoencoders trained both to reconstruct their inputs and to make their collective latent codes resemble a chosen probability...
• Adversarial Domain Adaptation - Adversarial domain adaptation is a family of methods that trains a useful predictive representation while making it harder to identify which domain supplied...
• Adversarial example - An adversarial example is an input deliberately modified to cause a machine-learning model to produce an incorrect prediction or other unintended behavior.

Sources

NIST AIRC, glossary context and terminology resource. https://airc.nist.gov/glossary/ River documentation, concept drift and drift-aware learning. https://riverml.xyz/latest/introduction/getting-started/concept-drift-detection/ River documentation, progressive validation and delayed labels. https://riverml.xyz/latest/api/evaluate/progressive-val-score/ Scikit-learn documentation, incremental learning and partial_fit. https://scikit-learn.org/1.8/computing/scaling_strategies.html The duration calculation and Willow Repairs case are original teaching illustrations.

Previous: Adaptive Huffman coding Next: Adaptive Learning Rate Methods