What is a neural network?
A neural network is a model of many connected computing units in layers, which adjusts its connections while it is trained.
The idea is loosely inspired by how nerve cells in the brain connect. In practice it is mathematics. Each connection has a weight, a number that decides how much it affects the next layer.
During training the weights are adjusted a little at a time, until the network gives the right answers on the examples. A large language model has a very large number of such weights.
That makes networks powerful but hard to inspect from the inside. You can rarely point to exactly why a particular answer turned out as it did.
How it works
A network has a layer that takes in the input, one or more hidden layers and a layer that gives the answer. Each unit receives numbers from the previous layer, weighs them, adds them up and passes the result on. Many simple steps like this, one after another, add up to an answer.
Why it matters
Neural networks are the foundation of deep learning and so of language models, speech recognition and image analysis. Understanding the principle explains why AI learns from examples, why it can be wrong and why the data it learned from matters so much.
What to keep in mind
A network learns every pattern in the data, including ones nobody intended. It can do well in testing yet worse on cases that differ from its training. Try it on real examples from your own work before you rely on it.
A network knows nothing beyond what is in its weights. It is unaware of anything that happened after training, unless it can read current documents when the question is asked.
What it means for you
Because the model is hard to inspect from the inside, its answers need to be checkable from the outside, for example with sources.
Common questions
Does a neural network work like a brain?
The name comes from a simplified picture of nerve cells. In practice it is mathematics, with numbers multiplied and added up in layers.
What does it mean that the network is trained?
It is shown examples, compares its output with the right answer and adjusts its connections a little at a time. After many rounds, the answers improve.
Related terms
Read more
All terms Updated