What Is a Neural Network (NN)? How Neural Networks Work, Explained Simply

A neural network (NN) is a type of machine learning model loosely inspired by the brain. It is made of layers of simple connected units called neurons. Each connection has a weight (a number), and by adjusting millions or billions of these weights during training, the network learns to turn inputs (pixels, words, sounds, numbers) into useful outputs (labels, predictions, text). Neural networks with many layers are called deep learning, and they power almost all modern AI, from image recognition to ChatGPT and Claude.

Neural networks can sound intimidating, but the core idea is simple. This guide explains how a neural network works step by step, how it learns, the main types and what each is used for, real-world examples, and the limitations to keep in mind.

The building blocks

Neurons (nodes)

An artificial neuron:

1. Receives several inputs (numbers)
2. Multiplies each input by a weight (how important that input is)
3. Adds them up, plus a bias (an adjustable offset)
4. Passes the result through an activation function that decides how strongly the neuron "fires"
5. Sends its output to neurons in the next layer

Layers

• Input layer: receives the raw data (e.g., pixel values of an image, numerical features of a customer)
• Hidden layers: transform the data step by step, learning increasingly abstract features
• Output layer: produces the result (e.g., probability that an image is a cat; the next word)

A network with many hidden layers is a deep neural network, which is where the term deep learning comes from.

Weights and biases

These are the network's parameters: the numbers it learns. Small networks have thousands; large language models have billions or more.

Activation functions

Non-linear functions (such as ReLU, sigmoid, tanh, GELU) that let networks learn complex, non-linear patterns. Without them, stacking layers would be no more powerful than a single layer.

A simple example: recognising handwritten digits

1. Input: a 28×28-pixel image of a handwritten digit = 784 numbers.
2. First hidden layer might learn to detect simple strokes and edges.
3. Next layers combine strokes into loops and lines.
4. Output layer has 10 neurons (digits 0–9); the one with the highest value is the prediction.

Nobody programs the "detect a loop" rule. The network discovers useful features by itself during training.

How a neural network learns (training)

1. Forward pass: feed an example through the network to get a prediction.
2. Loss calculation: compare the prediction with the correct answer using a loss function (how wrong was it?).
3. Backpropagation: calculate how much each weight contributed to the error, working backwards through the layers.
4. Gradient descent: adjust every weight slightly in the direction that reduces the error.
5. Repeat over many examples and many passes (epochs) through the data.

Over time, the weights settle into values that produce accurate predictions on data the network hasn't seen, provided it has enough good data and doesn't simply memorise the training set (overfitting).

Training large networks requires specialised hardware (GPUs and TPUs) because it involves enormous numbers of multiplications.

(For the full model-building process, see How is AI created?)

Types of neural networks

Where neural networks are used

• Language: chatbots and assistants, translation, summarisation, writing (see What does AI stand for?)
• Vision: face unlock, photo search, medical image analysis, self-driving perception, manufacturing defect detection
• Speech: voice assistants, transcription, voice synthesis
• Recommendations: streaming, shopping and social feeds
• Finance: fraud detection, credit risk, document processing
• Science: protein structure prediction (AlphaFold), drug discovery, weather forecasting
• Generative AI: text, image, music and video generation (see Which task is a generative AI task?)

Neural networks vs the human brain

Neural networks are inspired by the brain but are much simpler. Biological neurons communicate with complex electrochemical signals, learn continuously and use very little energy. Artificial neurons are simple mathematical functions trained in separate phases with large amounts of data and energy. The analogy helps intuition but shouldn't be taken literally.

Neural networks vs other machine learning

Neural networks excel with unstructured data (images, audio, text) and very large datasets. For structured, tabular business data (customer records, transactions), simpler models such as gradient-boosted trees are often as accurate or better, faster to train and easier to explain. (See What is machine learning? and Classification in machine learning.)

Limitations

• Data hungry: typically need lots of high-quality data.
• Compute and energy intensive, especially large models.
• Black boxes: hard to explain why a specific prediction was made, which is a problem in regulated decisions.
• Bias: they learn biases present in training data.
• Brittleness: small, carefully crafted input changes (adversarial examples) can fool them.
• Overfitting: can memorise rather than generalise without proper techniques.
• Hallucinations (in generative models): confident but incorrect outputs.

A tiny neural network in Python

from tensorflow import keras

model = keras.Sequential([
    keras.layers.Input(shape=(784,)),             # 28x28 image flattened
    keras.layers.Dense(128, activation="relu"),   # hidden layer
    keras.layers.Dense(64, activation="relu"),    # hidden layer
    keras.layers.Dense(10, activation="softmax")  # 10 digit classes
])
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy",
              metrics=["accuracy"])

(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train, x_test = x_train.reshape(-1, 784) / 255, x_test.reshape(-1, 784) / 255
model.fit(x_train, y_train, epochs=5, validation_split=0.1)
print(model.evaluate(x_test, y_test))   # typically ~97-98% accuracy

Want to learn more? See How to learn about AI.

Frequently asked questions

What is a neural network in simple terms?

A neural network is a computer model made of layers of connected units (neurons) that learns patterns from examples by adjusting the strength of its connections, allowing it to recognise images, understand language, make predictions and generate content.

What does NN stand for?

NN stands for neural network. You may also see ANN (artificial neural network), CNN (convolutional neural network) and RNN (recurrent neural network).

How do neural networks learn?

They make predictions, measure how wrong they are with a loss function, use backpropagation to calculate how each weight contributed to the error, and adjust the weights with gradient descent, repeating this over many examples.

What is the difference between a neural network and deep learning?

Deep learning refers to neural networks with many layers (deep neural networks). All deep learning uses neural networks, but a small neural network with one hidden layer isn't usually called deep learning.

Is ChatGPT a neural network?

Yes. ChatGPT, Claude and Gemini are built on very large neural networks using the transformer architecture, with billions of parameters trained on vast amounts of text and other data.

What are the main types of neural networks?

Feedforward networks (MLPs), convolutional neural networks (CNNs) for images, recurrent networks and LSTMs for sequences, transformers for language and more, autoencoders, GANs, diffusion models and graph neural networks.