In plain words
An artificial neuron is a very small calculation. It takes in a few numbers, gives each one an importance, called a weight, adds the weighted numbers together, and then decides what to send on.
Some weights are positive: that input pushes toward yes. Some are negative: that input pushes toward no. Big weights matter a lot, small ones barely count. The final step, a simple rule called an activation function, turns the total into the neuron's output.
One neuron on its own can only make a very simple decision. Link thousands or billions of them in layers and you get a neural network, able to recognise faces or write text. Training is the process of finding good weights for every neuron.
Everyday example
Deciding whether to go out for chai
Three things are on your mind. It is raining: a strong reason to stay in. Your friends are free: a stronger reason to go. Work is over: a small reason to go. You weigh them without thinking, the yeses outweigh the no, and you grab your umbrella. That weighing and adding is what one artificial neuron does.
Where the example breaks: You weigh reasons using everything you know and feel. An artificial neuron only multiplies numbers by fixed weights and adds them; its weights come from training, and on its own it understands nothing.
How it actually works
Despite the biology-inspired name, an artificial neuron is a short formula, and it is only loosely like a brain cell.
Weighted sum
Each input is multiplied by its weight and the results are added, often with an extra constant called a bias:
sum = w1×x1 + w2×x2 + w3×x3 + b. Google's glossary example: inputs 2, −1 and 3 with weights −1.3, 0.6 and 0.4 give −2.0.Activation
The sum goes through an activation function. The earliest neurons used a hard threshold: fire 1 above a line, 0 below. Today ReLU (negatives become 0) and smooth variants are common. A sigmoid turns Google's −2.0 into about 0.12.
Layers of neurons
In a network, each neuron's output becomes an input to neurons in the next layer. Stacked layers let simple units combine into complex detectors, which is the core of deep learning.
Learning the weights
The weights and biases are the network's parameters. Backpropagation and gradient descent adjust them, millions of times, to reduce errors.
Brain cells, loosely
Real neurons also collect signals and fire past a threshold, which inspired the idea. But they are far more complex, chemical and timing-based. Simple artificial neurons are sometimes called caricatures of the real thing.
- First described
- 1943, by Warren McCulloch and Walter Pitts, as a threshold logic unit
- Early learning machine
- Frank Rosenblatt's perceptron: a 1957 Cornell report, then a 1958 paper
- Also called
- Neuron, node, unit, perceptron (for the classic form)
- How many
- Large models contain millions to billions of them; their connection weights are counted as parameters
Questions people ask
What is an artificial neuron in simple words?
A tiny calculator inside a neural network. It takes some numbers in, multiplies each by a weight that says how much it matters, adds them up and passes the result through a simple rule. Its output feeds the next neurons.
How is an artificial neuron different from a real neuron?
A real neuron is a living cell that communicates with electrical pulses and chemicals and changes in complex ways. An artificial neuron is a few lines of maths. The idea of adding up signals and firing was borrowed from biology; little else was.
What is a perceptron?
An early artificial neuron that could learn its own weights from examples, built by Frank Rosenblatt in the late 1950s. It was the first famous learning machine. The word is still used for a single neuron or a simple layer of them.
Are neurons the same as parameters?
No. A neuron is a unit that does a calculation. Parameters are the numbers it uses, its weights and bias. One neuron can have hundreds or thousands of weights, one for each input.