An artificial neuron is the basic computing unit of an artificial neural network. It takes several numerical inputs, multiplies each by a weight that reflects its importance, adds the results together with a bias and passes the total through a function that decides what to output. Billions of these small units, wired together in layers, make up the systems behind image recognition, language models and many other forms of deep learning. The design was inspired by the nerve cells of the brain, but the artificial version is a simple mathematical simplification of the real thing.
Inspiration from biology
A biological neuron receives signals through branching fibers called dendrites, combines them in the cell body and, if the combined signal is strong enough, fires an electrical pulse along a long fiber called the axon to other neurons. The strength of the connections, the synapses, changes with experience, which is how learning is thought to occur. An artificial neuron borrows this outline. Inputs play the role of incoming signals, weights stand for synaptic strengths and the output stands for firing. Real neurons are far more complex, with chemical signaling, timing and many cell types, so the artificial neuron is a loose analogy and not a model of the brain.
Structure of an artificial neuron
An artificial neuron has four parts.
- Inputs: numbers x1, x2 and so on, which may be raw data or the outputs of other neurons.
- Weights: a number for each input, which scales its influence. A large positive weight amplifies an input, a negative weight reduces it and a weight near zero ignores it.
- Bias: a constant added to the sum, which shifts the point at which the neuron activates, much like the intercept of a line.
- Activation function: a function applied to the sum to produce the output, which introduces nonlinearity.
How an artificial neuron computes
The calculation has two steps. First, the neuron forms a weighted sum, multiplying each input by its weight, adding the products and adding the bias. Second, it applies the activation function to this sum. In symbols, the output is the activation function of the sum of weights times inputs plus the bias. For example, with inputs two and three, weights of one half and negative one, and a bias of one, the weighted sum is one minus three plus one, which equals negative one. A rectified linear unit, which outputs zero for negative values, would then give an output of zero.
Activation functions
Without an activation function, a neuron would be a linear function, and a stack of such neurons would collapse into one linear model that could not learn complex patterns. The activation function adds nonlinearity. Early neurons used a step function, which outputs zero or one. Later networks used the sigmoid, which squashes values between zero and one, and the hyperbolic tangent, which ranges from negative one to one. Modern networks mostly use the rectified linear unit, or ReLU, which is simple, fast and helps avoid the vanishing gradient problem, along with variants such as leaky ReLU, GELU and softmax for outputs that represent probabilities.
History of the artificial neuron
In 1943, the neurophysiologist Warren McCulloch and the logician Walter Pitts proposed a simple threshold model of a neuron and showed that networks of such units could compute logical functions. In 1949, Donald Hebb described how connections might strengthen through repeated use, an idea summarized as cells that fire together wire together. In 1958, Frank Rosenblatt introduced the perceptron, a neuron with learnable weights, and built a machine, the Mark I Perceptron, to recognize simple images. In 1969, Marvin Minsky and Seymour Papert showed that a single perceptron cannot solve some problems, such as the exclusive or function, which dampened enthusiasm. Interest revived in the 1980s when networks with hidden layers were trained with backpropagation, and it has grown ever since.
How neurons learn
Learning means adjusting the weights and biases so that the network’s outputs get closer to the desired ones. A loss function measures the error. The backpropagation algorithm calculates how much each weight contributed to the error, and gradient descent changes each weight a small step in the direction that lowers the loss. Repeating this over many examples gradually tunes the neurons so that they respond to useful patterns. Early layers of an image network, for instance, learn to detect edges, and later ones combine them into shapes and objects.
From one neuron to a network
A single neuron can draw only a straight dividing line between classes, which limits what it can do. Arranging neurons into layers overcomes this. Each neuron in a layer receives the outputs of the previous layer, and the layers together can represent very complex functions. The universal approximation theorem shows that a network with enough hidden neurons can approximate any continuous function to a chosen accuracy. Deep networks with many layers learn hierarchies of features, which is the basis of deep learning.
Variants of the artificial neuron
Researchers have created several variants. The perceptron uses a step function and a simple update rule. Sigmoid neurons give smooth outputs and are used in logistic regression. Convolutional neurons share weights across positions in an image. Recurrent neurons feed their output back to themselves to keep a memory of earlier inputs, and gated units such as LSTM and GRU control that memory. Spiking neurons, used in neuromorphic computing, communicate with discrete pulses over time and are closer to biology, and are the subject of research on energy-efficient hardware.
Applications and role
Artificial neurons are the building blocks of every neural network, so they appear in image classification, speech recognition, translation, recommendation, medical imaging, game-playing and large language models, which contain billions of weights. In practice, engineers rarely think about single neurons. They work with layers and architectures, but the behavior of the whole system emerges from the simple computation described above, repeated at enormous scale.
Limitations
A single neuron is limited to linear decisions. Large networks need much data and computing power, their internal workings are hard to interpret and they can learn biases in the data. The simple unit also omits much of biology, such as timing, chemistry and structural change, so claims that artificial neural networks work just like brains are overstated.
A worked example
Consider a neuron that decides whether to recommend an umbrella. It has two inputs, the chance of rain from zero to one and the wind speed scaled from zero to one. Training gives the rain input a weight of four, the wind input a weight of one and a bias of negative two. If the chance of rain is point eight and wind is point three, the weighted sum is three point two plus point three minus two, which equals one point five. A sigmoid activation turns that into about point eight two, which the app reads as a strong recommendation to carry an umbrella. Changing the weights changes how much each input matters, and learning is the process of finding good values.
Why the artificial neuron still matters
Despite its simplicity, the artificial neuron remains the foundation of modern AI. Every improvement in hardware, data and training methods acts on networks built from this one idea, a weighted sum followed by a nonlinear function. Understanding it makes larger concepts, such as layers, backpropagation and attention, much easier to follow.
FAQs
Is an artificial neuron the same as a perceptron? A perceptron is a specific kind of artificial neuron that uses a step activation function and a particular learning rule.
Why is a bias needed? It lets the neuron shift its activation threshold, so it can fit patterns that do not pass through zero.
How many neurons are in a large language model? Models have billions of parameters, which are weights and biases connecting many thousands of neurons in each layer.
More in Machine Learning
Key facts
- An artificial neuron computes a weighted sum of its inputs, adds a bias and applies an activation function to produce an output.1
- Warren McCulloch and Walter Pitts proposed a mathematical model of a neuron in 1943, and Frank Rosenblatt built the perceptron in 1958.2
- Networks of many artificial neurons arranged in layers form neural networks and deep learning models.1
Related concepts
Quick checkWhat does an artificial neuron compute?Show answer
A weighted sum of its inputs plus a bias, passed through an activation function to give its output.
Sources
- Stanford CS231n, Neural Networks Part 1: Setting up the Architecture, https://cs231n.github.io/neural-networks-1/
- McCulloch and Pitts, A logical calculus of the ideas immanent in nervous activity, https://link.springer.com/article/10.1007/BF02478259
- Wikipedia, Artificial neuron, https://en.wikipedia.org/wiki/Artificial_neuron
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
