D1 · The Perceptron
It Always Stops, If It Can
Say exactly what the perceptron convergence theorem promises, and what it refuses to promise.
Most learning algorithms come with a hope: run it long enough and it will probably get good. The perceptron comes with a proof. Under one condition, it is guaranteed to stop making mistakes after a finite number of nudges. Not probably. Guaranteed. The condition is the whole story.