Rosenblatt's perceptron: the machine that learned (1958)
Frank Rosenblatt's perceptron was the first machine to learn from experience, celebrated by the press, attacked by critics, and vindicated decades later.
In July 1958, the U.S. Office of Naval Research unveiled a machine that could learn. Its creator, Frank Rosenblatt, a research psychologist at the Cornell Aeronautical Laboratory in Buffalo, New York, called it “the first machine which is capable of having an original idea.” The press ran with the story. The New York Times reported on a “Navy device” that “learns by doing.”
The machine was the Mark I Perceptron, and it would become one of the most celebrated, and later most controversial, inventions in the history of artificial intelligence.
What the perceptron was
At its core, Rosenblatt’s perceptron was a single-layer neural network: a set of input units (simulating a retina of photocells), connected through adjustable weights to an output unit that produced a binary decision, yes or no, left or right.
The key insight was learning through adjustment. When the perceptron made an error, the weights on its connections were updated to make the correct answer more likely next time. Given enough training examples, it converged: the perceptron convergence theorem, proved by 1960, guaranteed that if the task was learnable by the architecture, training would find the right weights in finite steps.
Rosenblatt demonstrated this with a simple experiment: the Mark I could distinguish sheets of paper marked on the left side from those marked on the right. Modest by today’s standards, but in 1958, a machine that improved its own performance without being explicitly programmed was genuinely new.
Ambition and backlash
Rosenblatt was not shy about the implications. In his 1957 funding proposal, he wrote that devices of this sort were “expected ultimately to be capable of concept formation, language translation, collation of military intelligence, and the solution of problems through inductive logic.”
Not everyone was convinced. Marvin Minsky, who had known Rosenblatt since high school at the Bronx High School of Science, was deeply sceptical. The two debated publicly at conferences, loud, passionate, and personal, while their colleagues watched in amazement.
In 1969, Minsky and Seymour Papert published Perceptrons, a mathematical analysis that proved single-layer perceptrons could not learn certain functions, including the simple logical XOR. The theorems were correct. But the message that reached funding agencies and the broader community was blunter: “perceptrons don’t work.”
The nuance, that the limitations applied to single-layer networks, and that multi-layer networks might overcome them, was lost. Neural network funding dried up. The first AI winter (article) had arrived, at least for connectionist research.
A tragic end, a lasting legacy
Frank Rosenblatt died in a boating accident on Chesapeake Bay on July 11, 1971, at the age of 43. He did not live to see the vindication of his ideas.
That vindication came in the 1980s, when the backpropagation algorithm made multi-layer networks trainable, exactly the step Minsky and Papert had doubted was practical. The perceptron’s core principle, learning by adjusting weights from errors, is still the foundation of every modern deep learning system.
The Mark I Perceptron itself resides at the Smithsonian National Museum of American History. In 2004, the IEEE established the Frank Rosenblatt Award in his honour. The machine that learned turned out to be right, just sixty years too soon.