What is a perceptron?

One weighted decision

Suppose an input is represented by features x₁, x₂, …, xₙ. A perceptron gives each feature a weight, adds a bias, and calculates a score:

score = w₁x₁ + w₂x₂ + … + wₙxₙ + b

The same linear-threshold mechanism explains both panels: it can learn the blue separator when one straight boundary exists, but no single straight line separates opposite-corner XOR classes.

The sign of that score selects one of two classes. A positive weight makes its feature push the decision toward one class; a negative weight pushes toward the other. The bias moves the threshold without changing the feature values.

Geometrically, the weights and bias define a line in two dimensions, a plane in three, or a hyperplane in more dimensions. The output says which side contains the input. Calling the perceptron a “neuron” is a mathematical and historical analogy, not a claim that it reproduces a biological neuron.

Move the line, keep the four inputs

Use two binary inputs, x₁ and x₂. For OR, the positive class contains every point except (0, 0). Fixed weights 1, 1 and bias −0.5 separate it: the scores are −0.5, 0.5, 0.5, 1.5.

For XOR, only (0, 1) and (1, 0) are positive. The input coordinates stay identical; the required labels change. Predict what moving one straight boundary can fix—and which point will become wrong elsewhere.

Adjust the weights and bias below, then switch from OR to XOR. The classifications and mistake count are calculated locally from the four visible points.

A remaining XOR error is not evidence that the learning rate needs more tuning. One affine threshold cannot express this label pattern. The next article adds a nonlinear hidden representation rather than asking the same line to do more work.

Learning by correcting mistakes

Training presents labelled examples one at a time. If the perceptron predicts an example correctly, its weights can remain unchanged. If it predicts the wrong class, the update shifts the weights toward the correct label. Repeated mistakes therefore rotate and move the boundary.

For linearly separable training data, the classic convergence result says this procedure will eventually find a separating boundary under its assumptions. It does not promise the unique or best boundary, calibrated probabilities, or good performance on unseen data. Feature scale, ordering, margin, noise, and the relationship between training and production data still matter.

The XOR limit

Some class patterns cannot be divided by one straight boundary. XOR is the compact example: opposite corners of a square share a label, so every line that separates one pair splits the other pair incorrectly. A single perceptron cannot solve it regardless of how long it is trained.

That failure is architectural, not merely an optimizer problem. Combining units in layers and inserting nonlinear activation functions lets a network build intermediate representations whose final classes can be separated. This is the bridge from one linear threshold to a multilayer neural network.

Why engineers still learn it

Modern language models are vastly larger and use different components, losses, and training systems, but the perceptron exposes enduring ideas in a small form: numeric features, parameters, a forward calculation, a decision boundary, an update driven by error, and a capacity limit imposed by architecture.

It also prevents a common misconception. A learned model is not intelligent because a unit resembles a neuron or because training found weights. What it can represent follows from the complete architecture and data; what a deployed AI system can do also depends on the software around the model.