What Is an Artificial Neuron?

A single artificial neuron is almost embarrassingly simple. It does three things: it receives numbers, it multiplies each number by another number called a weight, and it adds everything up to produce a result. That is it. And yet, when you stack thousands or millions of these simple units together and train them carefully, they can recognize handwritten digits, detect tumors in X-rays, and translate between dozens of languages. The power is not in any individual neuron — it is in the organization. Before you can understand the organization, you need to understand the unit. Let's build it up from scratch.

Inputs: The Data Coming In

Every artificial neuron receives one or more input values. These are plain numbers — not words, not images directly, not sounds. Everything that enters a neural network must first be converted into numbers. A pixel's brightness becomes a number between 0 and 1. A word becomes a list of hundreds of numbers called an embedding. A sensor reading becomes a decimal value. Suppose you are building a very simple system to help decide whether a strawberry is ripe enough to pick. You might feed the neuron three inputs: the redness of the berry on a scale from 0 to 1, the firmness on a scale from 0 to 1, and the size relative to average, also from 0 to 1. Those three numbers go into the neuron simultaneously. The neuron has no idea what a strawberry is — it just receives three numbers.

Now for the interesting part: the neuron does not treat every input equally. Each input is multiplied by a weight before being added to the others. A weight is a number that controls how much influence that input has on the result. Back to the strawberry. Suppose redness gets a weight of 0.8, firmness gets a weight of 0.5, and size gets a weight of 0.2. Those weights represent a judgment that redness matters a lot (high weight), firmness matters moderately, and size matters less. When a berry is very red (0.9), moderately firm (0.6), and average-sized (1.0), the neuron computes: (0.9 × 0.8) + (0.6 × 0.5) + (1.0 × 0.2) = 0.72 + 0.30 + 0.20 = 1.22 That sum — 1.22 — is the neuron's raw output. We will talk about what happens to it next.

Bias: The Neuron's Own Nudge

There is one more ingredient that real neurons use: a bias. A bias is an extra number added to the weighted sum before producing the output. You can think of it as the neuron's default 'lean' before it even sees the inputs. If a bias is positive, the neuron is easier to activate; if negative, harder. So the full calculation is: (input_1 × weight_1) + (input_2 × weight_2) + ... + bias = raw output. In the strawberry example, if the bias is −0.5, then the raw output becomes 1.22 − 0.5 = 0.72. The bias shifts where the neuron 'sits' on its output scale. Every real neuron in every real network has both weights and a bias.

After computing the weighted sum plus bias, the neuron passes that number through one more step called an activation function. The activation function decides the final output. The simplest version just checks: is the number above zero? If yes, output something; if no, output zero. More sophisticated activation functions produce smooth curves that let the network learn more complex patterns. You do not need to memorize the math yet — the key point is that the activation function is what makes the network able to learn curved, non-linear relationships instead of only straight lines.