artificial neuron
/ ar-tuh-FISH-ul NOOR-on /
An artificial neuron is the tiny decision-maker that neural networks are built from. Picture a single voter weighing several pieces of news before deciding yes or no: each piece of news matters more or less to them, they have a personal leaning one way, and once they've added everything up they make a call. An artificial neuron does exactly that with numbers. It takes in several inputs, gives each one an importance score called a weight, adds them all up, adds a fixed nudge called a bias, and then passes the total through a small "shaping" function to produce one output number.
More precisely: if the inputs are numbers and each has a weight, the neuron computes the weighted sum (input times weight, all added together), adds the bias, and feeds that single number into an activation function. The activation function is what lets the neuron do more than plain arithmetic — it can squash the result into a range, or zero it out if it's negative. The output then becomes an input to neurons in the next layer. That is the whole machine: a weighted sum, plus a bias, run through one function.
The idea is loosely inspired by brain cells, which fire when enough signal arrives — but the resemblance is a metaphor, not a copy. A real neuron is staggeringly more complex. What makes the artificial version powerful is not any single neuron, which can only draw a straight dividing line, but wiring thousands or millions of them together so the network as a whole can represent rich, curved patterns. Alone, a neuron is almost trivial; in a crowd, they become a flexible learner.
A neuron deciding whether to recommend a movie sees two inputs: "is it a comedy?" (1) and "is it over two hours?" (1). It weights comedy at +2 and length at -1, and adds a bias of -0.5. The sum is (1×2) + (1×-1) + (-0.5) = 0.5. A ReLU activation passes anything positive through unchanged, so the neuron outputs 0.5 — a mild thumbs-up.
One neuron: weight each input, add them up, add a bias, then apply an activation.
A single neuron can only separate data with a straight line (or flat plane). Its power comes entirely from being stacked and combined with many others — and from the activation function, without which a whole network of neurons would collapse back into one straight line.