One of the first artificial neurons that could learn: a simple unit that adds up weighted inputs and answers yes or no depending on whether the total passes a threshold. Frank Rosenblatt introduced it in the late 1950s, and it learned by nudging its weights after each mistake. On its own it can only separate things with a straight line, a limit that once dampened enthusiasm for neural networks until multi-layer versions overcame it.
Technical view
Computes a weighted sum plus bias and applies a step function. The learning rule adds or subtracts each misclassified input from the weights, and the perceptron convergence theorem guarantees a finite number of mistakes when the data are linearly separable; otherwise it never settles. Rosenblatt described it in a 1957 report, publicly demonstrated a simulation on an IBM 704 in 1958, then built the Mark I Perceptron hardware for image recognition, shown in 1960. Minsky and Papert’s 1969 book Perceptrons formalised the limits of single-layer versions, including exclusive-or, and contributed to reduced funding. Multi-layer networks trained by backpropagation removed the limitation, and modern feed-forward blocks descend from it.
Related
Part of AI Terminology Index, section: How the Machinery Works.