In plain words
Picture many small decision-makers wired together in rows. Each one takes in a few numbers, decides how much to trust each, adds them up, and passes a single number on to the next row.
Those trust levels are called weights. A big weight means "this input matters a lot". A weight near zero means "ignore it". The first row receives the raw input, such as the words of an email. The last row gives the answer, such as "spam" or "not spam".
At first the weights are random and the answers are nonsense. During training, every weight is nudged a tiny bit, again and again, until the answers come out right. That is all learning means here.
The name comes from the brain, but the resemblance is loose. Each artificial neuron is just multiplication and addition. Stack many layers and you get deep learning.
Everyday example
A family deciding what to order
Should we order biryani tonight? Mum checks the budget, your brother checks the match timing, you check what you ate yesterday. Each opinion counts for a different amount, and Mum's counts most. Add them up and the family decides. A neuron does exactly this with numbers, and a network is many such little votes feeding into the next round.
Where the example breaks: A family can explain its votes. A trained network's weights are just numbers, and why one ended up at its value is usually hard to explain, which is the problem explainability research tackles.
How it actually works
Under the hood, a neural network is a long chain of multiply, add and bend.
The artificial neuron
Each neuron multiplies every input by its weight, adds them, adds one more number called a bias, and passes the total through an activation function that bends straight lines into curves. Without that bend, many layers would act like one.
Layers
Neurons are grouped into an input layer, one or more hidden layers and an output layer. Each layer's outputs become the next layer's inputs.
Weights are the knowledge
Weights and biases together are the network's parameters. A small network may have thousands. A large language model has billions.
Learning by backpropagation
The output is compared with the right answer using a loss function. Backpropagation traces the error back through every layer to find how each weight should change, and gradient descent makes the change.
- First maths model of a neuron
- 1943, Warren McCulloch and Walter Pitts
- An early learning machine
- The perceptron, by Frank Rosenblatt at the Cornell Aeronautical Laboratory, shown in July 1958
- Its famous demo
- After 50 tries it learned to tell cards marked on the left from cards marked on the right
- Today
- Every modern chatbot and image model is a very large neural network
Different wiring patterns make different families: CNNs for images, RNNs for sequences, transformers for language.
Questions people ask
Is a neural network like the human brain?
Only loosely. It borrows the idea of many simple units passing signals through connections of different strength. Real neurons are far more complex and use chemistry and timing. Think of an artificial network as inspired by the brain, not a copy of it.
What is the difference between a neural network and deep learning?
A neural network is the structure. Deep learning is the practice of training neural networks with many layers. A network with one small hidden layer is still a neural network, but is not usually called deep.
What are weights in a neural network?
Weights are the numbers on the connections between neurons. They decide how much one neuron's output affects the next. Training is the search for good values. Together with biases they make up the model's parameters.
Why are neural networks called black boxes?
Because their knowledge is spread across millions of unlabelled weights. We can see every number but cannot easily read why the network gave a particular answer. Explainability research tries to open the box.