Neural Networks Explained Plainly

The basic building block: the "neuron"

An artificial neuron takes several numbers as inputs, multiplies each by a "weight" that says how important that input is, adds the results together with a small adjustable number called a bias, and passes the total through a simple function that decides how strongly to "fire". It is loosely inspired by biological neurons, though the real mechanism is quite different.

A tiny worked example: two inputs, 0.8 and 0.2, with weights 0.9 and -0.5 and a bias of 0.1.
0.8 × 0.9 = 0.72, and 0.2 × -0.5 = -0.10. Adding them gives 0.62, and adding the bias gives 0.72.
A common activation function, ReLU, passes positive values through unchanged and turns negative values into zero, so this neuron outputs 0.72.

That is all a neuron does. The power comes from having millions of them wired together.

Layers stack these neurons together

A network arranges neurons into layers. Data flows in through an input layer, passes through one or more hidden layers where patterns are combined, and a result comes out at the output layer.

Simple example: in an image-recognition network, early layers may detect edges, middle layers combine edges into features like an eye or an ear, and the final layer combines those features to decide "cat".

Where the weights come from: training

At the start the weights are random, so the output is nonsense. Training adjusts them gradually using a process called backpropagation. The network makes a guess, compares it with the right answer, works out how wrong it was and which weights contributed most to the error, and nudges those weights slightly. This repeats millions of times (see How AI Actually Learns).

Why "deep" learning is called deep

"Deep" just means many hidden layers between input and output. Modern large language models have dozens of layers, and some have over a hundred.

The architecture behind today's chatbots

Chatbots such as ChatGPT, Claude and Gemini are built on a design called the transformer. Its key idea, attention, lets the model weigh how much each token in the text matters to every other token, so that in a sentence like "the trophy didn't fit in the suitcase because it was too big" it can work out what "it" refers to. You do not need the maths to use these tools, but it helps to know that the model reads text as tokens and relates them to each other through attention.

Why layered networks work so well

Each layer builds on the one before it, turning simple low-level patterns into increasingly abstract ones. That layered composition lets neural networks handle tasks too complex for hand-written rules.

Frequently asked

Do I need to understand the math to use AI effectively?

No — using AI tools well doesn't require understanding the underlying mechanics. This lesson matters more if you're building or fine-tuning models yourself.

Are neural networks the only kind of machine learning?

No — simpler techniques like decision trees and linear regression exist and are sometimes better suited to smaller, more structured problems, as covered in an earlier lesson.

How many "neurons" does a large language model have?

Modern large models can have hundreds of billions of parameters (the weights between neurons) — vastly more than any earlier generation of neural networks.

← Back to Technical Track Next: Training vs Inference →