The Perceptron
The original artificial neuron (Rosenblatt, 1958) — and it really learns, right here.
What is it?
A perceptron takes numeric inputs, multiplies each by a weight, adds a bias, and outputs 1 if the sum is positive, else 0. That's it. It's a linear classifier: in 2D, its decision rule is literally a straight line.
Watch it learn — live
This is a real perceptron training in your browser, not a canned animation. Click to add points (toggle class with the buttons), drag existing points around, and press Train. The line moves every time the perceptron makes a mistake and corrects its weights: w ← w + lr·(target − prediction)·x.
The learning rule
For each training point the perceptron predicts, then compares to the true label. If it's right, nothing happens. If it's wrong, the weights are nudged toward misclassified class-1 points and away from misclassified class-0 points:
b ← b + lr · (target − prediction)
The learning rate controls the size of each nudge. The perceptron convergence theorem guarantees that if the data is linearly separable, this procedure finds a separating line in finitely many updates.
The famous limitation: XOR
Press Load XOR above. Four points in an XOR pattern cannot be split by any single straight line — so the perceptron thrashes forever. Minsky & Papert pointed this out in 1969, contributing to the first "AI winter." The fix — stacking neurons into layers with non-linear activations — is exactly what the multi-layer perceptron does. That's your next stop.