Learn AI /Neural Networks /Neural network
Neural network
- Starter
- Neural Networks
- Policy officers
- Technical readers
Also called: artificial neural network, neural net, ANN.
Back to the milestone you came from
In five seconds
What it is
A neural network is a layered set of connected numerical operations that turns input data into an output such as a label, score or prediction.
In plain words
A neural network takes input data, passes it through several layers of calculations, and produces an output such as a label or score.
An everyday example
Imagine several sound mixers connected in a row. On each mixer, every input passes through a slider that makes it stronger or weaker. The mixer combines the adjusted inputs, adds a small baseline setting and uses a gate to decide what signal to pass onward. The next mixer repeats the process with the signals it receives. During rehearsal, the sliders and baseline settings are adjusted whenever the final sound differs from the target. After many corrections, the same settings can turn a new set of inputs into an output. A neural network similarly learns weights and biases across layers, although it processes numbers rather than sound.
A real neural network does not contain people who understand language or meaning. Its internal steps are fast numerical transformations, usually all applied mathematically at once, and training changes numbers in the system rather than teaching separate human-like specialists.
Remember thisA neural network is a chain of learned numerical transformations that produces a label, score or prediction. Its output reflects its training data and parameters, not human judgement.
Understand it
Understand it
Why it matters
Non-specialists often need to judge claims about AI systems. Understanding neural networks helps you ask whether a system has enough data, whether its outputs are only scores rather than facts, whether human oversight is needed, and whether the task truly requires this approach.
How a neural network turns input into output
After viewing, the reader should be able to say that a neural network passes numbers through layers of weighted calculations to produce an output.
01
1. Input numbers enter
The system starts with numeric representations of the data, such as word counts, pixel values or sensor readings.
Render a left column titled Input layer with 4 circles containing example values like 0.8, 0.1, 0.6, 0.3. Use arrows pointing right.
02
2. Hidden layer combines them
Each unit multiplies incoming values by weights, adds a bias, and computes a new value.
Render a middle column titled Hidden layer with 3 circles. On hover or tap, show a small label: weighted sum + bias, then activation.
03
3. Signals continue forward
The new values become inputs to the next layer. Stronger and weaker pathways affect the final pattern.
Use lines of varying thickness from input to hidden and from hidden to output to suggest different weights. Animate a pulse moving left to right if motion is allowed.
04
4. Output scores appear
The final layer returns scores or probabilities for the task, such as likely categories.
Render a right column titled Output layer with 2 circles labelled for example damaged road: 0.82 and no damage: 0.18. Highlight the larger value.
05
5. Training changes connections
When the answer is known during training, the system adjusts weights and biases to reduce future error.
Add a feedback arrow from output to connection labels below the diagram reading adjust weights during training, styled differently from the forward arrows.
Show three columns of circles labelled input, hidden and output. Animated arrows carry example numbers left to right. Each hidden circle displays “weighted sum + bias”, then an activation step. The final output shows two scores, with the larger one highlighted. A small note below says training adjusts weights after comparing outputs with known answers.
Hovering or tapping a circle reveals a short tooltip with its current value and role. A play button runs one example input through the network; a reset button returns to the initial state.
Describe this without the diagram
This diagram explains a neural network as a left-to-right flow. First, an input layer receives numbers representing the original data. Second, a hidden layer combines those numbers using weighted sums and bias terms, then applies an activation function to each result. Third, those new values move forward to the output layer. Fourth, the output layer shows final scores for possible answers, with the highest score marked as the current result. Fifth, a separate feedback arrow notes that during training, known answers are used to adjust the weights so later outputs improve.
With reduced motion: A static three-column diagram replaces animation. Numbers are shown already placed in the input, hidden and output layers, and text labels explain that values move left to right during a forward pass while a separate feedback note indicates that training later adjusts weights.
A neural network is a machine-learning method built from connected numerical operations, often organised in layers. The terms “artificial neural network”, “neural net” and “ANN” usually mean the same broad idea. “Neural network” is the common short form, while “artificial” reminds us that this is an engineered system, not a biological brain.
What goes in depends on the task. The input might be numbers from a spreadsheet, pixels from an image, or words turned into numbers. In a basic feed-forward network, a unit multiplies inputs by adjustable values called weights, adds a bias term, and passes the result through an activation function. Other neural-network designs organise and combine numerical operations differently.
What happens when an input is processed is often called a forward pass. Values move through the network to produce an output. In many training tasks, the network compares its output with a target and adjusts its weights and biases to reduce error over many examples.
What comes out might be a category, a probability-like score, a translated phrase, or a forecast. A neural network does not contain facts in the way a database does. It produces outputs by applying learned numerical patterns, so results can be useful but still wrong, biased or overconfident.
People often confuse neural networks with deep learning. Deep learning usually means neural networks with many layers and large-scale training, so it is a subset, not a separate unrelated idea. Specific architectures such as convolutional neural networks, recurrent neural networks and transformers are different designs within the wider neural-network family.
How “A row of adjustable sound mixers” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the report moving through several people in order | input values moving through layers |
| each person highlighting some parts more than others | weights giving different importance to inputs |
| marked-up notes passed onward | intermediate numerical activations |
| the final English summary | the network output |
A worked example
A toy neural network assigns a match score to a fictional application using only simplified, job-related features. The example illustrates the calculation, not an acceptable recruitment method.
Starting from: Synthetic features: required certification present = 1, relevant experience match = 0.8, required language evidence = 1, domain-term match = 0.2.
- The features enter the first layer as the numerical vector [1, 0.8, 1, 0.2].
- Each hidden unit multiplies the inputs by its learned weights, adds a learned bias and applies an activation function. One unit might calculate 0.6 × 1 + 0.3 × 0.8 + 0.2 × 1 − 0.1 = 0.94, which remains 0.94 after a rectified linear activation.
- Other hidden units use different weights, so they respond to different combinations of the same features. Their activations become the inputs to the next layer.
- The output unit combines those activations and applies a function that produces a score between 0 and 1, for example 0.81.
- During training, the network compares its outputs with labelled examples and adjusts its weights and biases to reduce a loss function. Once trained, those parameters produce the score for a new input.
Result: The output is an illustrative match score of 0.81. It is a learned prediction, not a finding that the fictional applicant is eligible or suitable.
In a United Nations workplace
A duty officer in a specialised agency’s technical division is covering the recruitment focal point and must complete a screening matrix and note for the file before the applications go to the chief of section at close of business. A vendor proposes a neural network that assigns each application a match score against the roster profile.
The score is produced by successive layers of learned numerical operations, not by the system applying the profile as a panel would. The duty officer needs to establish which features enter the model, how its weights were trained, what the score means, where errors occur and whether the result can be reviewed and challenged.
Before you try thisApplications contain personal data and may reveal sensitive characteristics. They must not be uploaded to a public or otherwise unapproved AI tool. Any approved use requires authorised access, appropriate retention controls, testing for discriminatory effects and meaningful human review. A neural-network score must not be the sole basis for excluding an applicant.
Where you may hear it
- A vendor demo during a procurement meeting
- An ICT governance board reviewing a business case
- A front-office briefing before clearance of an AI pilot
- An evaluation meeting examining model errors
Avoid confusion
Avoid confusion
Often confused with
Deep learning
A method that trains many-layered computer models to find useful patterns in large amounts of data.
Transformer
A neural network design that processes many parts of a sequence in parallel by weighting which parts matter most to each other.
Brain analogy (and its limits)
A comparison between neural networks and brains that can help with intuition, but breaks down because AI systems use simplified maths rather than living cells and biological processes.
Common misconceptions
Common misconception: A neural network works like a human brain.
Correction: The name comes from a loose inspiration, but the mechanism is a set of mathematical operations on numbers. Real brains are far more complex, and the analogy can mislead if taken literally.
Common misconception: A bigger neural network is always better.
Correction: More layers or parameters can improve performance on some tasks, but they also raise costs, data needs and failure risks. For many problems, a simpler model may be more suitable and easier to test.
Common misconception: If the output score is high, the answer must be true.
Correction: The score reflects how strongly the model matches learned patterns, not a guarantee of correctness. Models can be confidently wrong, especially on unfamiliar or biased data.
Common misconception: Neural network and deep learning mean exactly the same thing.
Correction: Deep learning usually refers to neural networks with many layers and large-scale training. It is a major part of neural-network use, but not every neural network is deep.
Quick check
Which statement best describes a neural network?
The correct option describes the broad neural-network family. Deep learning uses many-layered models, a transformer is one neural-network design, and the brain comparison does not mean that neural networks reproduce living cells.
Go deeper
Go deeper
The simple version
A neural network takes input data, passes it through several layers of calculations, and produces an output such as a label or score.
The technical version
An artificial neural network is a parameterised function composed of interconnected units arranged in layers. Each unit computes a weighted sum of inputs plus a bias, applies an activation function, and passes the result forward; training updates parameters to minimise a loss function over examples.
Reviewed external resources
EncyclopaediaNeural network (opens in a new tab)Wikipedia
This reference page gives a broad overview of definitions, history and related concepts, and suits readers who want a general starting point.
Intermediatelink checked 2026-07-31
InteractiveTensorFlow Playground (opens in a new tab)TensorFlow
This interactive tool lets you change layers, weights and data and see effects directly, and suits learners who understand best by experimenting.
Starterlink checked 2026-07-31
VideoBut what is a neural network? (opens in a new tab)3Blue1Brown
This video gives an intuitive visual introduction to neural networks and suits beginners who want the big picture before technical detail.
Starterlink checked 2026-07-31
VideoNeural networks, explained visually (opens in a new tab)3Blue1Brown
This video offers another visual explanation of how signals move through layers, and suits visual learners who want a mechanism-focused view.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Neural network (opens in a new tab)YouTube search
This is a YouTube search results page, not a single vetted explanation, and it suits readers who want to compare several video styles carefully.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Neural network (opens in a new tab)Reddit search
This is a Reddit search results page, not a single authoritative source, and it suits readers who want to see common questions, opinions and misunderstandings discussed by users.
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
A neural network is a chain of learned numerical transformations that produces a label, score or prediction. Its output reflects its training data and parameters, not human judgement.