Key takeaways
- A neural network passes input through layers of units joined by weighted connections.
- Training adjusts the weights from examples so the network's answers improve.
- Connectionism is the name of the research tradition that builds intelligence from such networks.
- Because knowledge is spread across many weights, people cannot read a network's reasoning directly.
How a neural network is built
A neural network is built from an input layer, one or more hidden layers and an output layer, each made of simple units.
Each unit takes numbers from the units before it, multiplies each by a weight, adds them up and passes a result on. The input layer receives the data, such as the words of an email turned into numbers. The output layer gives the answer, such as how likely the email is to be an automatic reply. The hidden layers in between are where the network builds its own intermediate features.
How do neural networks learn
Neural networks learn by comparing their output with the right answer and adjusting every weight a little to shrink the error.
The method most networks use is called backpropagation: the error at the output is traced back through the layers to work out how much each weight contributed, and each weight is nudged in the direction that helps. Repeated over many examples, these small changes add up to a network that generalises to new input. Nobody writes the rules; they emerge from the data.
Connectionism and the black box
Connectionism is the tradition in AI and cognitive science that explains intelligence as the product of many simple connected units, and its main trade-off is that the result is hard to read.
A symbolic system stores a rule you can read. A neural network stores what it learns as a very large number of weights, none of which means anything on its own. That is the root of the black box worry: the network may be right, but it cannot show a chain of reasons a person can check, and its errors can be hard to predict. Explainable AI research works on this problem.
Neural network vs symbolic AI
A neural network learns its behaviour from examples, while symbolic AI follows rules people write.
Neural networks do well on messy, varied input such as language, images and sound, where writing rules by hand is hopeless. Symbolic AI does well where steps must be exact, explained and the same every time. Neurosymbolic research tries to get both by pairing the two.
A neural network example in sales
An illustrative example is a network that learns to tell a genuine out-of-office auto-reply from a person saying they are not interested.
After seeing thousands of labelled replies, the network picks up the cues on its own: phrasing, return dates, signatures, the way people word a refusal. No one wrote a rule for each case, so it copes with wording it has not seen before. It will also misfile some replies with no obvious reason, so a team that acts on its labels still checks the ones that matter, such as a refusal that also names a better contact.
Where Pebbles Ai fits
Common questions
Are neural networks modelled on the human brain?
Only loosely. The units and weights borrow the idea of neurons and their connections, but real brains work very differently, so the comparison is best treated as a metaphor.
What is an artificial neural network?
It is the full name for a neural network in computing, used to separate it from the biological networks in the brain. In AI writing the word artificial is usually dropped.
What is the difference between a neural network and deep learning?
Deep learning is the use of neural networks with many hidden layers. Every deep learning model is a neural network, but a small network with one hidden layer is not usually called deep.
About this page. Written by the Pebbles Ai team for people who run sales, marketing and growth. Last updated .
