1958: the first thing that LEARNED. A single artificial neuron adjusts its weights every time it guesses wrong, nudging its decision line toward the mistake — and on data that CAN be split by a line, it’s guaranteed to find one, in finite time. It’s the seed of every neural network. (Its famous limit — it can’t learn XOR — is why we stack layers.) Slide the epochs and watch the line snap into place.
The perceptron is a linear binary classifier: y = sign(w·x + b). It learns by the PERCEPTRON RULE — on each misclassified example, w ← w + y·x (and b ← b + y) — nudging the boundary toward the error. The Perceptron Convergence Theorem guarantees it finds a separating hyperplane in finitely many updates IF the data is linearly separable. It cannot solve non-linearly-separable problems (famously XOR, Minsky–Papert 1969), which motivated multi-layer networks and backprop. A fail-loud self-check throws unless it reaches 0 errors on linearly-separable data. ◆ real machine learning, node-verified.
Convergence is guaranteed ONLY for linearly-separable data (else it never settles); the single neuron is the ancestor of, not a substitute for, deep networks. The update rule and convergence are exact.