The Perceptron

Editorial record

Plain-language summary

In 1958 Frank Rosenblatt described the perceptron, a simple network that combines weighted inputs and produces a decision, and a procedure that updates those weights whenever it makes a mistake. This gave a working example of a machine that improves its performance by seeing labeled data rather than being explicitly programmed. The perceptron could learn to separate patterns that are linearly separable, though later work showed its limits on harder problems. Its learning rule and layered structure are the foundation of today's deep learning.