The History of AI

The perceptron

Era 2 · AI is born (1943–1979) › Neural networks and learning

Lineage
Neurons and networks
Kind
Implementation

Who

Cornell Aeronautical Laboratory, Office of Naval Research

What happened

Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, describes the perceptron in a 1957 report and a 1958 paper in Psychological Review. In July 1958 the US Navy shows a simulation on an IBM 704, and in 1960 the Mark I Perceptron machine is unveiled.

Why it mattered

The first network that learns its weights from examples, with a rule that corrects the connections every time it gets something wrong.

The full story

Rosenblatt started from the McCulloch-Pitts neuron and Hebb's rule but asked a different question: not how an already built network computes, but how a network can organize itself from experience. The perceptron had a layer of sensors, like a retina, randomly connected to intermediate units, which in turn connected to response units whose weights were adjusted. If the response was wrong, the weights were corrected in the direction that would have given the right one.

In July 1958 the Office of Naval Research showed the press a simulation on an IBM 704: after about fifty trials, the system learned to tell cards marked on the left from cards marked on the right. The New York Times reported that a new Navy device "learns by doing," and Rosenblatt spoke of a machine able to perceive and recognize its surroundings without human control. Then came the Mark I Perceptron, a visual classifier with a retina of 400 photocells and weights stored in potentiometers turned by electric motors. It was assembled at the Cornell Aeronautical Laboratory in Buffalo in 1959, unveiled in public on June 23, 1960, and is now held by the Smithsonian.

Only the last layer learned: the first had fixed random connections. Rosenblatt proved that if a problem can be solved with a single adjustable layer, the rule will find the solution, but there was no method for training several layers, though by 1962 Rosenblatt was already writing about propagating errors backward.

The promises in the press and the limits Minsky and Papert pointed out in 1969 shaped the perceptron's reputation. Rosenblatt died in 1971, and the idea revived with backpropagation in the 1980s.

Three inputs, x1, x2, and x3, are joined by lines to a node that sums them. Each line carries a weight, w1, w2, and w3. The sum passes through a step-shaped threshold and comes out as 0 or 1.

A perceptron multiplies each input by its weight, adds up the results, and outputs 1 if the sum exceeds the threshold.

Under the hood

The response unit computes y = 1 if w₁x₁ + w₂x₂ + … + wₙxₙ + b > 0, and y = 0 otherwise. The weights w and the bias b are what is learned.

The learning rule acts only when the response is wrong: it adds its input to each weight if the unit should have output 1, and subtracts it if the unit should have output 0. If it is right, nothing changes. Rosenblatt proved that if some set of weights classifies all the examples correctly, the rule finds it after a finite number of corrections.

In the Mark I only that last layer was adjusted. The connections from the photocell retina to the intermediate units were random and fixed.

Builds on

  • The McCulloch-Pitts neuron (December 1943)

    Documented direct influence

    Rosenblatt starts from the McCulloch-Pitts threshold neuron and discusses their model in the 1958 paper

  • Hebb's rule (1949)

    Documented direct influence

    The 1958 paper cites Hebb and takes from Hebb the idea that connections change with activity

Leads to

  • Perceptrons (1969)

    Documented direct influence

    The book analyzes the perceptron mathematically and is named after it

Threads that pass through here

Sources

How to cite

“The perceptron.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/rosenblatt-perceptron

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline