14 SEP 2026 · 09:10 · FOUNDATIONS
What Is a Neural Network? A Practical Explanation
A neural network is a trainable mathematical system that learns useful input-to-output mappings by adjusting many connected numerical weights.
A neural network is a machine-learning model made from layers of connected calculation units. It receives numbers, transforms them repeatedly, and produces an output such as a class, a probability, a forecast or the next piece of text. The word “neural” is historical inspiration, not a claim that the software thinks like a brain.
The short answer
The useful mental model is a large adjustable function. During training, examples show the system which outputs are better. An optimisation process changes the network’s weights so that, across many examples, its predictions become less wrong. During inference, those learned weights are held fixed while a new input passes through the network.
Input, hidden and output layers
The input layer represents the features given to the model. For an image, those may begin as pixel values; for a tabular prediction, they may be measurements; for language, text is first converted into tokens and then numerical vectors.
Hidden layers sit between the input and the output. Google’s Machine Learning Crash Course describes the units in hidden layers as neurons. Each unit combines incoming values using learned weights and a bias, then passes the result through an activation function. Stacking these transformations lets a network represent relationships that a single straight-line model cannot.
The output layer is shaped for the task. An image classifier may output a probability for each category. A regression model may output one number. A language model outputs a probability distribution over possible next tokens, then repeats the process as text is generated.
What weights and activation functions do
A weight controls how strongly one signal influences the next calculation. A positive weight can reinforce a signal, a negative weight can suppress it, and a value near zero can make it matter very little. A bias shifts the point at which a unit responds.
Without activation functions, many layers of weighted sums would collapse into another linear calculation. Activations introduce non-linearity, which is what allows a network to model curved boundaries, interactions and other complex patterns. The exact activation matters, but the core idea is simple: it decides how the combined signal is passed forward.
How a neural network learns
Training usually starts with weights that do not yet encode a useful solution. The model makes a prediction, a loss function measures the error, and backpropagation calculates how much each weight contributed to that error. An optimiser then nudges the weights in a direction expected to reduce the loss. Repeating this process over batches of examples gradually fits the network to the training data.
Learning is not the same as memorising every example, although memorisation can occur. The goal is generalisation: useful performance on new inputs drawn from the intended setting. A validation set helps tune choices without repeatedly looking at the final test set, while an untouched test set provides a more honest estimate of performance.
A small example
Suppose a network predicts whether an email needs urgent attention. Inputs might encode the sender, time, message length and representations of the words. Hidden layers can learn combinations such as an unusual sender plus deadline language plus a near-term date. The output might be a score between zero and one.
That score is not an explanation or a guarantee. It is the result of the patterns learned from the training examples and the way the system was evaluated. If urgent messages in the real inbox differ from those examples, accuracy can fall even when the training score looked strong.
Why deep networks became important
A deep neural network has more than one hidden layer. Depth gives a model room to build representations in stages. In image systems, early layers may respond to local edges while later layers combine them into larger patterns. In language systems, successive transformations can connect words with their context and construct task-relevant representations.
More layers and parameters do not automatically make a model better. Data quality, objective design, optimisation, compute, regularisation and evaluation all matter. A smaller model trained and tested for the real task can be more useful than a larger one chosen from a headline benchmark.
Where neural networks fail
Neural networks can inherit bias from data, become overconfident on unfamiliar inputs, and exploit shortcuts that correlate with the target without representing the intended concept. Their internal representations are distributed across many weights, which makes a complete human-readable explanation difficult.
The practical response is not to treat the model as magic. Define the task, keep a representative evaluation set, test important subgroups and failure cases, monitor changes after deployment, and decide what level of human review the consequence demands. A neural network is a powerful fitted function; its reliability remains conditional on the evidence around it.
Sources and further reading
- Google for Developers — Neural networks: nodes and hidden layers
- Google for Developers — Machine Learning Glossary