What Is an Artificial Neuron? Weights, Bias, Activation

Signals in, one number out

Real neurons in your brain collect electrical signals and fire when enough arrive. An artificial neuron is a loose, simplified cousin of that idea. It does not copy biology in detail. It is a small recipe:

1. take a handful of input numbers;
2. scale each one by its own weight;
3. add the scaled values together and add a bias;
4. pass the total through an activation function.

The result is one number. You can read it as a score: how strongly does this neuron believe the thing it is watching for is present? If the inputs are the brightness of two sensors, the output might say "looks like daytime".

Weights: how much each input matters

A weight is a dial that scales one input before it joins the sum. A big positive weight means "this input counts a lot in favour". A weight near zero means "mostly ignore this input". A negative weight means "this input pushes against".

Signs combine. A positive input times a negative weight pulls the total down, and so does a negative input times a positive weight. A negative input times a negative weight pushes it up. Because every input has its own weight, a neuron can learn to care about some signals and disregard others.

In maths terms, the weighted sum is w₁x₁ + w₂x₂ + … + wₙxₙ. If you have met vectors, this is a dot product: the weight vector against the input vector.

Bias: the neuron's starting mood

The bias is a single extra number added after the weighted sum. It shifts the neuron's output up or down no matter what the inputs are. It sets how readily the neuron says "yes".

With a large negative bias, the inputs must add up to a lot before the neuron responds. With a large positive bias, it responds easily, even to weak inputs. Without a bias, a neuron's decision would always have to pass through zero, which is far too rigid for real data.

Altogether the neuron computes output = activation(w₁x₁ + w₂x₂ + b).

Activation: squashing the total

If the neuron just reported the raw sum, a stack of neurons could be collapsed into one straight-line calculation, however many you added. The activation function bends the result and breaks that limit.

Common choices include:

• step: output 0 or 1, a hard yes or no;
• sigmoid: a smooth S-shape that squeezes any total into a value between 0 and 1;
• ReLU: output zero for negative totals and pass positive totals through unchanged.

Different activations suit different jobs, but all of them play the same role: they turn a plain sum into something a network can build on.

From one neuron to a decision line

With two inputs, a single neuron makes a surprisingly geometric decision. Plot every possible input pair as a point on a plane. The neuron says "class A" when w₁x₁ + w₂x₂ + b is positive and "class B" when it is negative. The boundary between them, where the sum is exactly zero, is a straight line.

Changing the weights tilts that line. Changing the bias slides it. So learning, at this scale, means finding the weights and bias that place the line where it separates your two groups best. That is all training does; larger networks simply have many more dials and bend the lines into curves.

A lone neuron can only separate data with a straight line. That is its limit, and it is the reason networks stack layers of them.

Common misconceptions

• "A neuron understands things." It multiplies, adds and squashes. Meaning comes from the pattern across many of them.
• "Weights are fixed rules." They start as guesses and are adjusted during training.
• "Bias means unfairness here." In this context it is only a shift value. (Biased training data is a separate, real concern.)
• "More neurons always means smarter." Size helps only with good data and training.

How to explore this in the lab

The AI Lab lets you change a neuron's dials and watch the output and decision line respond in 3D. Work through the free first world in order:

1. Meet the parts in What is a neuron?.
2. Turn each dial in Weights, bias, activation.
3. Draw a boundary in From a neuron to a decision.

The lab is a simplified teaching model with tiny examples. Real networks have thousands to billions of weights, but each neuron in them does what you see here.