Glossary
What is an artificial neuron?
In short
An artificial neuron is a simple calculation unit of a neural network: it multiplies input numbers by weights, adds them and passes the result on through an activation function.
AI basics · artificial neuron
MMOOZZGG.com editorsChecked: 2 min read
For a child
Imagine a line of kids who pass hints to each other in a whisper. Each one listens to the neighbors, decides how important the hint is and passes it on louder or quieter. One such player knows almost nothing. But when there are very many of them standing in rows, together they solve a hard riddle. In a computer neural network, such a player is called a neuron.
In plain words
An artificial neuron was inspired by the nerve cell of the brain, but it is much simpler. It gets several numbers, and each input has its own weight, that is, how important it is. The neuron multiplies the inputs by the weights, adds them up and decides how strongly to “light up” and what to send to the next units. Imagine you are choosing a film: a friend’s review weighs a lot for you, an ad very little, and you decide by the total whether to go. A neural network has a huge number of such units, and training changes exactly their weights.
For business
A manager does not need to understand the formulas, but it helps to know what follows from them. The number of neurons and the connections between them set the size of the model: the bigger it is, the more expensive the server and each answer. The model’s knowledge is spread across the weights of a huge number of connections, so you cannot open it and fix one rule, as in a spreadsheet. Errors are fixed with data: fine-tuning, examples in the prompt, checking the answer at the output. This is also why decisions are hard to explain: to the question “why was the customer refused” the network gives no direct answer. Where a company needs such a model, and where simple rules are enough, is clear from the diagram of the AI floors: data, system links, analysis, action.
For a specialist
An artificial neuron computes y = f(w·x + b): the dot product of the input vector x and the weight vector w, a bias b and a nonlinear activation function f. W. McCulloch and W. Pitts proposed the first formal model in 1943 (Bulletin of Mathematical Biophysics): the unit fires on an “all or nothing” principle when a certain number of synapses are excited, and an active inhibitory synapse fully blocks firing. Thanks to the “all or nothing” principle, networks of such units can be described by propositional logic. In 1958, F. Rosenblatt described the perceptron, a network of similar units that learns from experience. Modern networks replaced the hard threshold with smooth or piecewise linear functions: sigmoid, hyperbolic tangent, ReLU, GELU. This is needed to compute gradients in backpropagation. Neurons are grouped into layers, and a fully connected layer is implemented as the product of a weight matrix and an input vector.
Example
A model that estimates the risk of a customer leaving gets as input the number of calls to support, the time since the last purchase and the order value. One neuron in the first layer can learn to react to the combination “many complaints and no purchase for a long time”, and the next layers combine such signals into the final score.
Related terms
Questions and answers
Sources
- W. S. McCulloch, W. Pitts, A logical calculus of the ideas immanent in nervous activity, Bulletin of Mathematical Biophysics 5, 1943link.springer.com
- F. Rosenblatt, The perceptron: a probabilistic model for information storage and organization in the brain, Psychological Review 65(6), 1958homepages.math.uic.edu
Not sure whether your company needs this?
AI solutions map