A neural network is a type of machine learning model made up of layers of interconnected nodes, loosely inspired by neurons in the human brain. By adjusting the strength of connections during training, a neural network learns to recognise patterns in images, speech, and text that would be impossible to program with explicit rules.
What is a biological neuron and how does it inspire the artificial version?
A biological neuron receives electrical signals from many other neurons through connections called synapses. If the combined signal is strong enough to pass a threshold, the neuron "fires" — sending its own signal onwards. The strength of each synapse changes with experience: this is how the brain learns.
An artificial neuron mimics this:
| Biological neuron | Artificial neuron |
|---|---|
| Receives signals from other neurons | Receives numerical inputs |
| Synapse strength | Weight (a number that scales each input) |
| Fires if threshold exceeded | Applies an activation function to the weighted sum |
| Sends signal to next neurons | Outputs a number to the next layer |
The key insight: a single artificial neuron is trivially simple, but connecting millions of them in layers produces a system capable of extraordinary pattern recognition.
What are the layers of a neural network?
A standard neural network has three types of layer:
Input layer → Hidden layers → Output layer
[Pixel 1] \ [Cat: 92%]
[Pixel 2] → [Hidden] [Hidden] → [Dog: 7%]
[Pixel 3] / [Bird: 1%]
...
Input layer — one node per input feature. For an image of 28×28 pixels, there are 784 input nodes, each receiving a pixel brightness value between 0 and 1.
Hidden layers — one or more intermediate layers where patterns are extracted. Early layers might detect edges; later layers might detect shapes; final layers might detect specific objects. Networks with many hidden layers are called deep neural networks, which is where the term deep learning comes from.
Output layer — one node per possible answer. An image classifier trained on ten animal types has ten output nodes; the node with the highest value is the network's prediction.
How does a neural network learn?
Learning in a neural network means adjusting the weights on the connections between nodes until the network's outputs match the correct answers in the training data. This is done through a process called backpropagation:
- Forward pass — feed an input through the network; compute the output.
- Calculate error — compare the output with the correct answer using a loss function (a measure of how wrong the network was).
- Backward pass — work backwards through the network, adjusting each weight to reduce the error. Weights on connections that contributed to the mistake are reduced; those that helped get the right answer are reinforced.
- Repeat — process thousands or millions of training examples, each time nudging the weights slightly closer to correct.
After enough training examples, the network generalises — it can classify images, transcribe speech, or translate text that it has never seen before.
What can neural networks do?
| Application | What the network does | Input | Output |
|---|---|---|---|
| Image recognition | Classifies what is in a photo | Pixel values | Label ("cat", "car", …) |
| Speech-to-text | Transcribes spoken words | Audio waveform | Text |
| Language translation | Converts between languages | Text in source language | Text in target language |
| Spam detection | Classifies emails | Word frequencies | "Spam" or "Not spam" |
| Medical imaging | Detects tumours in scans | MRI/X-ray pixel data | Diagnosis probability |
| Game playing | Selects the best move | Board state | Move recommendation |
What are the limitations of neural networks?
Neural networks are powerful but have important weaknesses:
Black box — even their designers cannot always explain why a network makes a particular decision. This "explainability problem" is serious when the decision affects someone's medical diagnosis or loan application.
Data hungry — a neural network for image recognition may need millions of labelled training images. Collecting and labelling this data is expensive and time-consuming.
Biased by training data — if the training data is unrepresentative, the network will reproduce those biases. A face-recognition system trained predominantly on one demographic will perform worse on others.
Computationally expensive — training large networks requires enormous amounts of processor time and energy. A single large language model may use as much electricity during training as a small town uses in a week.
Frequently asked questions
What is a neural network in simple terms for KS3?
A neural network is a computer system inspired by the human brain. It consists of layers of connected nodes (artificial neurons). By training on thousands of examples, it learns to recognise patterns — such as which pixels in a photo make it a cat — and can then classify new examples it has never seen before.
What is the difference between machine learning and a neural network?
Machine learning is the broad field of computer systems that learn from data rather than from explicit rules. Neural networks are one particular type of machine learning model. Other machine learning approaches include decision trees, support vector machines, and naive Bayes classifiers. Deep neural networks (with many hidden layers) are especially powerful for image, audio, and language tasks.
Why are neural networks called "deep" learning?
"Deep" refers to the depth — the number of hidden layers — in the network. Early neural networks had one or two hidden layers. Modern networks can have dozens or hundreds of layers, allowing them to learn increasingly abstract representations. More depth (up to a point) allows the network to solve more complex tasks. "Deep learning" became prominent in the early 2010s when deep networks, combined with large datasets and powerful GPUs, began outperforming all other methods on image and speech recognition.
Are neural networks always the best approach?
No. Neural networks need large amounts of labelled training data and significant computing power. For smaller datasets, simpler machine learning methods (decision trees, linear regression) often work equally well or better, while being far easier to interpret. Neural networks shine when the input is high-dimensional (images, audio, natural language) and large training datasets are available.
Explore how artificial intelligence really works with Professor Turing at aitutors.me — probing questions to build your conceptual understanding.