Meet the Neuron
Loading simulation…
flagWhat you'll discover
- arrow_forwardDescribe an artificial neuron as weighted inputs plus an activation
- arrow_forwardExplain what weights and biases do
- arrow_forwardShow how a non-linear activation lets a neuron "fire"
- arrow_forwardPredict whether a neuron will output high or low for given inputs
A neuron is a tiny decision-maker
An artificial neuron is a simplified mathematical model of a brain cell, and it is the atom from which all neural networks are built. It takes several numbered inputs — think of them as signals arriving along wires — multiplies each by a number called a weight, and adds them all up along with a bias term. The weight says how important each input is; the bias shifts the whole score up or down.
That sum is then passed through an activation function, which squashes the result into a final output. If the weighted sum is high enough, the neuron "fires" and outputs a value near 1; if not, it stays quiet near 0. A single neuron is, in effect, a tiny decision-maker: "given these inputs, with these importances, should I switch on?"
Weights and bias
The weights are where the neuron's knowledge lives. Suppose a neuron decides "should I take an umbrella?" from two inputs: is it cloudy, and did the forecast say rain. If cloudiness matters less than the forecast, the forecast input gets a bigger weight. The neuron multiplies each input by its weight, adds them, and the bigger the total, the more it leans towards "yes, take one".
The bias is like the neuron's default mood — a number added to the sum before the activation. A high positive bias makes the neuron trigger easily ("paranoid, always carries an umbrella"); a strongly negative bias makes it reluctant. Tweaking the weights and bias is exactly what training does: learning is the slow adjustment of these numbers until the neuron gives the right answers.
Why activation matters
Without an activation function, a neuron would just multiply and add — and a whole network of pure adders could be collapsed back into one big linear equation, able only to draw straight boundaries. The activation function bends that line, letting the neuron respond non-linearly: small inputs are ignored, then suddenly the neuron snaps on.
The most popular activation today is called ReLU (Rectified Linear Unit): it outputs zero for any negative sum and the raw value for any positive one — dead simple, yet stacked in millions it produces everything from image recognition to ChatGPT. In the simulation, drag the input sliders, watch the weighted sum add up on the glowing wires, and see the output light up when it crosses the activation threshold.