Perceptron

The perceptron was introduced by psychologist Frank Rosenblatt in 1957 at the Cornell Aeronautical Laboratory. It was a single-layer, feed-forward neural network designed to recognise visual patterns — images of triangles, circles and squares — using an array of photocells connected to adjustable weights. The Mark I Perceptron, a room-sized machine built in 1960, was the first device to learn from examples rather than being explicitly programmed.

Rosenblatt proved that if a classification problem was linearly separable, the perceptron would converge to a correct solution. But the 1969 book Perceptrons by Marvin Minsky and Seymour Papert showed that a single-layer perceptron could not solve non-linear problems such as XOR. The critique was technically correct — for a single layer — but it was interpreted broadly as a condemnation of all neural networks. Funding dried up and AI entered its first 'winter'.

The field revived in the 1980s when researchers applied backpropagation to multi-layer perceptrons (MLPs). Hidden layers could learn non-linear representations, overcoming the limitations Minsky and Papert had identified. Today every deep neural network — from ResNet to GPT-5 — traces its lineage back to Rosenblatt's original idea: a trainable weighted sum followed by a non-linear activation.