What is Machine Learning (ML)? | AI Glossary

Start here

Machine Learning (ML) in plain language

A support team could write a rule that sends every message containing “invoice” to billing. A learned classifier instead uses examples to recognise a wider range of wording. Neither approach is automatically better: a simple rule can be easier to maintain when the task is predictable.

Supervised learning uses examples with target answers. Unsupervised learning looks for structure without those target labels, while reinforcement learning learns from rewards associated with actions. The data, objective and evaluation must match the intended task.

How it works

The visual guide

Learning is not the same as passing the test

Build and evaluate a supervised model

Training examples
Fit patterns using inputs and their target labels.
Validation examples
Compare settings and choose the candidate model.
Held-out test
Estimate performance on examples not used for those choices.

After deployment, monitor new cases. Do not repeatedly tune against the final test set.

Keep related records together when splitting data. For a forecasting task, preserve time order so future information cannot leak into training.

Define the target
Choose what should be predicted and when that prediction will be used. Include a simple baseline for comparison.
Prepare representative examples
Check labels, missing values and duplicated records. Separate development data from the final evaluation.
Measure the errors that matter
Inspect false positives and false negatives, not only overall accuracy. Check performance on the cases users actually encounter.

Illustrative example · Fictional data

Routing fictional support requests

Input: A set of labelled messages belongs to billing, access or delivery support. Some messages from the same conversation are near-duplicates.
Process: Keep each conversation in one data split. Train on the training portion, choose settings on validation data and compare with a keyword baseline.
Expected output: A suggested queue for each new message, with uncertain cases routed to a person.
How to check it: If duplicates appear in both training and testing, an impressive score can reflect memorisation. Also inspect messages that mention an invoice but actually report a login problem.

Where it helps

Classifying incoming documents or messages.
Forecasting demand from relevant historical patterns.
Finding unusual observations for further investigation.

Rules, machine learning and deep learning

Rules: People specify the decision logic. Useful when conditions are clear and stable.
Machine learning: Model parameters are learned from data. Useful when patterns are difficult to describe completely as rules.
Deep learning: A part of machine learning using multilayer neural networks, rather than an alternative to ML.

How much data and specialist support do you need?

There is no universal minimum dataset size. The requirement depends on task difficulty, label quality, class balance, variation and the performance target. A simple baseline on a carefully separated test set is more informative than collecting a large volume without checking whether it represents deployment conditions.

Transfer learning reuses representations from an existing model and adapts them to another task. It can reduce the amount of task-specific training needed, but the model still needs suitable data and evaluation. Test performance on relevant subgroups and on cases that differ from the training distribution.

A managed prediction API can suit a common task when its behaviour and terms fit. Custom development needs ownership of data preparation, evaluation, deployment and monitoring—not necessarily a large team, but explicit responsibilities. Compare a simple rule or statistical method before assuming a complex model is necessary.

Common pitfalls

More data can reinforce the wrong pattern if it is unrepresentative or incorrectly labelled.
Information unavailable at prediction time can leak into the training features.
A model’s performance can change when users, products or operating conditions change.

Put the idea to work

Try it yourself

Write ten fictional messages and assign a support queue to each. Create a keyword rule, then find two messages it misroutes. Describe what extra evidence a learned system would need.

Common questions

Does machine learning always need labelled data?

No. Supervised learning uses target labels, but other learning approaches use different signals or look for structure without labels. Unsupervised learning finds clusters or patterns, self-supervised learning creates labels from the data itself, and reinforcement learning learns from rewards. Large language models rely heavily on self-supervised pretraining.

Does a deployed model keep learning automatically?

Not necessarily. Many deployed models stay fixed until a separate update or retraining process runs. Teams usually monitor performance, collect new data and then retrain and re-evaluate the model on a schedule or when quality drops, rather than updating it continuously in production.

Is high accuracy enough?

No. Check the baseline, the evaluation data and the consequences of different errors. A rare but important case can be hidden by a high overall score. For example, a fraud model that is 99% accurate may miss most fraud if fraud cases are rare. Precision, recall and cost-weighted metrics give a fuller picture.

Your next step