What is a neural network in simple terms?

A neural network is a mathematical system loosely inspired by the brain, made of many simple units called neurons connected in layers. Each neuron takes numbers in, multiplies them by learned weights, adds them up, and passes the result through a function to produce an output. Stacked together, these simple units can learn to recognize images, translate text, and much more.

Is a neural network actually like a human brain?

Only loosely. The idea of interconnected neurons that fire based on their inputs is borrowed from biology, but artificial neurons are far simpler than real ones and learn through maths, not chemistry. The analogy is helpful for intuition, but a neural network is really a large, trainable mathematical function, not a digital brain.

What are weights and biases?

Weights are numbers that decide how much each input matters to a neuron — a large weight amplifies an input, a small one ignores it. The bias is an extra number added in that shifts the neuron's output up or down. Learning means adjusting all these weights and biases so the network's outputs match the correct answers.

Why do neural networks need so many layers?

Each layer transforms its inputs into a more useful representation, and stacking layers lets the network build up from simple features to complex ones — edges to shapes to objects in an image, for example. More layers (depth) let the network model very complicated relationships, which is why the field is called deep learning.

What Is a Neural Network? How AI Learns Like a Brain

The core idea

A neural network is a mathematical system, loosely inspired by the brain, built from many tiny units called neurons. Each neuron does something simple: it takes some numbers as input, multiplies each by a weight, adds the results together along with a bias, and passes the total through an activation function to produce a single output number. On its own, one neuron is almost trivial. The magic comes from connecting thousands or millions of them into layers, so that the outputs of one layer become the inputs of the next. Together they form a flexible function that can be shaped — trained — to map inputs like pixels or words to outputs like labels or translations.

A single neuron, step by step

Picture one neuron with two inputs. It computes a weighted sum: input one times weight one, plus input two times weight two, plus the bias. That single number then flows through an activation function such as the sigmoid, which squashes any value into the range 0 to 1. The weights decide how much each input matters, and the bias shifts the threshold at which the neuron “fires.” The playground above lets you set these by hand: move the weights and bias, change the inputs, and watch the output respond. This is the entire computation a neuron performs — everything else in a network is many copies of it.

From one neuron to many layers

Real networks stack neurons into layers. An input layer receives the raw data, one or more hidden layers transform it, and an output layer produces the final answer. Each layer learns to detect more abstract features than the last: in an image network, early layers might respond to edges, middle layers to shapes, and later layers to whole objects. Because each layer feeds the next, the network builds complex understanding out of simple steps. The term deep learning simply means a network with many such layers — depth is what lets these systems model very complicated relationships.

How a network produces an answer

Running a network forward is called the forward pass. Data enters the input layer, flows through each hidden layer — every neuron computing its weighted sum and activation — and emerges at the output layer as a prediction. For a digit recognizer, the output might be ten numbers giving the probability the image is each digit. Nothing about the network changes during a forward pass; it is just applying the weights it currently has. This is exactly what happens every time you use a trained model: a fast, fixed forward pass over learned weights.

Where learning comes in

So far the weights were assumed to be correct, but at the start they are random and the network’s answers are nonsense. Learning is the process of adjusting all those weights and biases so the outputs match known correct answers. The network compares its prediction to the truth, measures the error with a loss function, and then uses backpropagation and gradient descent to nudge every weight in the direction that reduces the error. Repeat this over a large dataset and the random network gradually becomes a capable one. That training process is a topic in its own right — but it all rests on the simple neuron you can experiment with above.