An artificial neuron is the basic computing unit of a neural network. It takes several numeric inputs, multiplies each by a learned weight, adds them together with a bias, and passes the result through an activation function to produce one output. On its own, a neuron is simple arithmetic. Connected in layers by the thousands or billions, neurons form neural networks and deep learning models.
Key takeaways
- A neuron computes: output = activation(weighted sum of inputs + bias).
- Weights set how much each input matters; the bias shifts the decision point.
- The activation function decides how the neuron responds and adds non-linearity.
- A single neuron can only separate data with a straight line, which is why networks need many neurons in layers.
- It was loosely inspired by biological neurons but works very differently.
An artificial neuron explained in simple terms
Think of a neuron as a tiny voter deciding “yes” or “no”. It listens to several advisers (the inputs). It trusts some advisers more than others (the weights). It has its own starting mood, leaning towards yes or no before hearing anyone (the bias). It adds everything up and, if the total is convincing enough, it votes yes (the activation). Training is the process of adjusting how much the voter trusts each adviser, so its votes become more accurate.
The parts of an artificial neuron
| Part | Symbol | Role |
|---|---|---|
| Inputs | x₁, x₂ … xₙ | Numbers coming in, from raw data or from other neurons |
| Weights | w₁, w₂ … wₙ | Learned values that scale each input up, down or negative |
| Bias | b | A learned offset that makes the neuron easier or harder to activate |
| Weighted sum | z | z = w₁x₁ + w₂x₂ + … + wₙxₙ + b |
| Activation function | f | Transforms z into the output, for example ReLU or sigmoid |
| Output | y | y = f(z), sent to the next layer or used as the prediction |
A worked example: a neuron that computes AND
Let us build a neuron that outputs 1 only when both inputs are 1. Set weights w₁ = 1 and w₂ = 1, bias b = −1.5, and use a step activation (output 1 if z is above 0, otherwise 0).
| x₁ | x₂ | z = x₁ + x₂ − 1.5 | Output |
|---|---|---|---|
| 0 | 0 | −1.5 | 0 |
| 0 | 1 | −0.5 | 0 |
| 1 | 0 | −0.5 | 0 |
| 1 | 1 | 0.5 | 1 |
Change the bias to −0.5 and the same neuron computes OR instead. That shows the bias’s job: it moves the threshold. In real networks, nobody sets these values by hand. Training finds them automatically using backpropagation and gradient descent.
What the neuron is really doing
Geometrically, a single neuron draws a straight line (or a flat plane in more dimensions) through the data and outputs a different value on each side. That line is its decision boundary. The weights set the line’s angle and the bias sets its position. This is why a single neuron behaves much like logistic regression when it uses a sigmoid activation.
The limit of one neuron: the XOR problem
XOR outputs 1 when exactly one input is 1. Plot the four cases and you will see no single straight line can separate the 1s from the 0s. A single neuron therefore cannot learn XOR. Marvin Minsky and Seymour Papert highlighted this limit in their 1969 book Perceptrons, which contributed to a long slowdown in neural network research. The solution is to add a hidden layer: two neurons draw two lines, and a third combines them. Stacking neurons in layers is what gives networks their power.
Common activation functions
- Step: outputs 0 or 1; used in the original perceptron but not trainable with gradients.
- Sigmoid: smooth curve from 0 to 1; useful for probabilities.
- Tanh: smooth curve from −1 to 1.
- ReLU: outputs 0 for negative values and the value itself for positive ones; the default in most hidden layers.
- GELU and SiLU: smooth variants common in modern transformer models.
Artificial vs biological neurons
| Aspect | Biological neuron | Artificial neuron |
|---|---|---|
| Signals | Electrical spikes and chemicals, timed | Plain numbers, no timing |
| Connections | Thousands of synapses that change physically | Weights stored as numbers |
| Learning | Many biological mechanisms, not fully understood | Gradient-based updates from a loss function |
| Complexity | A single cell can perform complex computation | A multiply, add and one function |
The name reflects historical inspiration, not a close copy. Modern networks are engineering tools, not brain simulations.
A short history
In 1943, Warren McCulloch and Walter Pitts published a mathematical model of a neuron that fired when its inputs crossed a threshold. In 1958, Frank Rosenblatt introduced the perceptron, a neuron that could learn its weights from examples. After the XOR critique, progress slowed until backpropagation was popularised in 1986, allowing multi-layer networks of neurons to be trained together.
Artificial neuron vs perceptron vs node
These terms overlap. Neuron and node are used interchangeably for any unit in a network. A perceptron usually means Rosenblatt’s specific model with a step activation, though people also say “multilayer perceptron” for simple feedforward networks.
Limitations
A single neuron is linear in its decision boundary and cannot model complex patterns alone. Individual neurons in large networks are hard to interpret; many respond to mixtures of concepts. And with some activations, neurons can “die”: a ReLU neuron that always receives negative input outputs zero and stops learning.