Neural Networks Explained Simply

How do you teach a computer to spot a cat in a photo without giving it a rule for it? Modern AI’s answer is neural networks that learn it themselves from examples.

What a neural network is

A neural network consists of many small computing units arranged in layers. Each unit takes in numbers, combines them and passes on a new signal.

At the front the input goes in, such as the pixels of an image. At the back the output comes out, such as cat or dog. In between lie hidden layers that do the actual work.

The loose model of the brain

The idea comes from the brain, where nerve cells exchange signals over connections. Hence the name neural network.

But the similarity is only rough. An artificial network does not think and feels nothing. It is pure mathematics, a vast system of numbers and simple computing steps.

How training works

Every connection in the network has a weight, a small number. At first these weights are random, and the network only guesses.

During training it is shown many examples with known answers. When it is wrong, the weights are nudged a tiny bit in the right direction. After millions of passes, the network gives good results.

Deep networks and deep learning

Modern networks have very many layers. This is called deep networks or deep learning. Each layer recognizes somewhat more abstract features.

For an image, the first layer finds edges, a later one shapes, an even later one whole objects. This depth made AI’s great leap from around 2012 possible.

What they can and cannot do

Neural networks are strong at recognizing patterns. They process images, language and measurement data often better than any hand-written rule.

But they do not understand in the human sense. They only know patterns from their data and can clearly go wrong with novel cases or gaps. How powerful language models grow from this is shown in the artificial intelligence section.