Learn AI /How Machines Learn /Training
Training
- Starter
- How Machines Learn
- Policy officers
- Technical readers
Also called: training a model.
Back to the milestone you came from
In five seconds
What it is
Training is the process of adjusting a model's parameters using data or feedback to improve performance on a learning objective.
In plain words
Training means showing a model many examples and adjusting it so its future outputs are more useful.
An everyday example
At a water utility, a technician shows a new trainee many past sensor logs from pumps in different towns. On each log, they mark whether the pattern was normal operation or an unusual reading that later needed inspection. The trainee first makes many mistakes, so the technician points out what was missed, such as a sudden pressure drop combined with vibration spikes. After seeing many examples and corrections, the trainee becomes better at flagging new logs.
A human trainee can use broad background knowledge and may explain their reasoning. A trained model usually adjusts numeric parameters instead, and its internal changes are not the same as human learning or understanding.
Remember thisTraining adjusts a model’s parameters using examples and a learning objective. Good results on training data alone do not show that the model will generalise to new inputs.
Understand it
Understand it
Why it matters
If you hear that a system was trained, you should ask what data shaped it, what feedback it used, and whether it was tested on new cases. Those details affect accuracy, fairness, safety and whether the system fits the task at all.
How training changes a model
After viewing this, the reader should be able to say what goes into training, what changes during it, and what comes out for later use.
01
1. Provide examples
Training starts with a dataset of examples. Some methods also include target answers or other feedback.
Draw several rectangular cards entering the first box. Label a few cards with short sample values and, where appropriate, a target tag such as 'normal' or 'fault'.
02
2. Produce trial outputs
The current model processes the examples and produces initial predictions or representations.
Show the cards moving into a simplified model shape, with output tokens appearing on the right side of the box.
03
3. Measure the gap
A training procedure compares the trial output with the target or desired pattern and calculates how wrong it was.
Display two columns, 'model output' and 'target', with mismatched items highlighted and a meter labelled 'difference'.
04
4. Adjust and repeat
The model's internal settings are nudged to reduce the difference, and the cycle repeats many times until performance improves.
Show slider controls shifting slightly on the model icon, plus a loop arrow back to step 2 and a final badge saying 'trained model'.
A left-to-right step diagram with four connected boxes. Box 1 shows a stack of example cards labelled 'data'. Box 2 shows a model icon producing trial outputs. Box 3 shows a comparison panel between trial outputs and target answers, with a small gauge labelled 'difference'. Box 4 shows the same model icon with sliders moved and a label 'updated model for later use'. Arrows connect each box, and a curved arrow from Box 3 back to Box 2 shows repetition.
On hover or tap, each step reveals one sentence of extra detail and highlights the corresponding arrow in sequence; a replay control restarts the highlight loop.
Describe this without the diagram
This diagram explains training as a four-step process. First, a set of examples is provided, sometimes with correct answers or another form of feedback. Second, the current model produces trial outputs from those examples. Third, the training procedure compares those outputs with the target or desired pattern and measures the difference. Fourth, the model's internal settings are adjusted to reduce that difference, and the same cycle repeats many times. The end result is an updated model that can later be used on new inputs.
With reduced motion: A static four-box diagram shows the same process without animated arrows. A numbered loop marker beside steps 2 to 4 indicates that those steps repeat many times before the final updated model is produced.
A machine learning system does not usually start with fixed written rules for every case. Instead, it starts with a model, which is a mathematical structure with many adjustable settings, and a dataset, which is a collection of examples. Training uses those examples to change the settings.
What goes in depends on the kind of learning. In supervised learning, the input often includes examples plus the correct answers, often called labels. In other approaches, the system may look for structure in unlabelled data or learn from feedback signals. The important point is that training uses data to shape the model.
During training, the model makes trial outputs on the training data. A training procedure measures how far those outputs are from the target or desired pattern, then nudges the internal settings to reduce that difference. This happens many times, often over batches, which are small groups of examples processed together.
What comes out is not a list of remembered answers for every possible case. The result is an updated model, often saved as a checkpoint, that can later be used for inference, meaning applying the trained model to new inputs. Good training aims to help the model work well on new data, not only on the examples it saw before.
How “Learning to spot a suspicious pump reading” maps onto the real thing
| In the story | In the AI system |
|---|---|
| past sensor logs with outcomes | training data |
| marks showing normal or unusual | labels or feedback targets |
| the trainee's changing judgement | the model's adjusted internal parameters |
| repeated correction after mistakes | iterative parameter updates during training |
A worked example
A developer trains a simple model to classify short messages as a question, a request or feedback.
Starting from: An approved set of example messages, each assigned one of the three categories, divided into training and validation sets.
- The model begins with parameter values that do not yet capture useful patterns, so its initial classifications are unreliable.
- For each batch of training messages, it predicts a category and a loss function measures the difference between those predictions and the assigned categories.
- The training algorithm uses that loss to adjust the model’s parameters. Repeating this process over many batches gradually reduces the training error.
- The developer checks the updated model on the separate validation set. If performance there does not improve, further training may only make the model fit the training examples too closely.
Result: Training produces an updated model whose parameters reflect patterns found in the examples. It can then classify a new message, but validation and human review remain necessary because lower training error does not guarantee reliable performance on unseen text.
In a United Nations workplace
Following a bilateral on a proposed text-classification tool, a statistician in the data unit at a Vienna duty station prepares the note for the file before close of business. The supplier proposes training the model on previously coded survey responses in three working languages so that it can suggest categories for the next monitoring report.
The coded responses provide examples of the desired output. During training, the system processes them in batches, measures how far its predictions are from the assigned categories and adjusts its parameters to reduce that error. Testing on responses excluded from training indicates whether it can generalise beyond examples it has already seen.
Before you try thisRaw survey responses may contain personal or otherwise sensitive information. They must not be uploaded to a tool that has not been approved for such data. Use an approved environment, remove unnecessary identifiers, restrict access and retention, and test performance separately across the three languages. Suggested categories still require human review.
Where you may hear it
- A vendor demo during a procurement meeting
- A technical review before an ICT governance board
- An evaluation office discussion of model performance and bias
- An inter-agency working group reviewing data requirements
Avoid confusion
Avoid confusion
Often confused with
Inference
Using a trained model on new input to produce an output such as a score, category, translation or generated text.
Fine-tuning
Updating a pre-trained model's parameters with additional examples to improve performance on a narrower task, style or domain.
Prompt
The text, image or other input that guides a model on what kind of output to produce next.
Side-by-side
Common misconceptions
Common misconception: Training means typing instructions into a system.
Correction: A prompt gives directions at use time, but training changes the model itself using many examples or feedback signals over repeated updates.
Common misconception: A trained model has simply memorised the dataset.
Correction: Some memorisation can happen, especially with weak methods or poor data practices, but the goal of training is to learn patterns that also work on new data.
Common misconception: Once training finishes, the model stays reliable forever.
Correction: Performance can drop when the real world changes, the training data were limited, or the model was overfitted, so monitoring and review still matter.
Common misconception: More training data automatically means a better system.
Correction: Quality, relevance, labelling, representativeness and bias matter as much as quantity, and bad data can reinforce errors.
Quick check
Which statement best describes training in machine learning?
Training is the stage where the model's parameters are updated from data or feedback. Using the model afterward is inference, and prompting does not usually change the underlying model.
Go deeper
Go deeper
The simple version
Training means showing a model many examples and adjusting it so its future outputs are more useful.
The technical version
Training is an optimisation process that updates model parameters using data and a learning objective, often by minimising a loss function over repeated batches, to improve generalisation to unseen inputs.
Reviewed external resources
EncyclopaediaWikipedia search: Training (opens in a new tab)Wikipedia
This Wikipedia search results page helps readers scan several possible meanings of 'training' and suits someone who wants a broad starting point rather than one settled article.
search results, not one pagelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course gives a structured introduction to core machine learning ideas, including training workflows, and suits beginners who want guided lessons.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source is useful for step-by-step explanations with intuition and suits learners who prefer clear spoken walkthroughs.
Starterlink checked 2026-07-31
VideoVideo explanations: Training (opens in a new tab)YouTube search
This YouTube search results page lets readers compare different video explanations of training and suits people choosing a teaching style that fits them.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Training (opens in a new tab)Reddit search
This Reddit search results page shows community questions and informal explanations about training and suits readers who want to see common confusions and practical discussion.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Training adjusts a model’s parameters using examples and a learning objective. Good results on training data alone do not show that the model will generalise to new inputs.