← Back to timeline
Classical ML1958interactive demo live

Perceptron

One neuron, one boundary

A single weighted sum plus threshold showed that machines could learn a classifier directly from data.

Perceptron Demo

Click in the chart to drop a point of the selected class. Then run the perceptron — each pass updates the weights only for the points it gets wrong. The line below is w · x + b = 0.

live demo
-5-3-1135-5-3-1135

click chart to add a + point · misclassified points have a yellow ring

Current hyperplane

1.00·x₁ 1.00·x₂ + 0.50 = 0

Epoch
0
Errors
12
Acc
0%
Add a point with this label
Train
0.10.501.5

The update rule

For each misclassified point, push the weight vector in the direction of its true class:
w ← w + η · y · xb ← b + η · y. Correctly classified points are skipped — they pull no force.

Why it converges

If the classes are linearly separable, the perceptron is guaranteed to find a separating hyperplane in finite steps (Novikoff, 1962). Try a non-separable cloud — it will oscillate without settling.