Historical Context
The world in 1958
Eight years had passed since Turing’s paper. The world had changed fast.
The Soviet Union had launched Sputnik in October 1957 — the first satellite in orbit. America was in a state of scientific panic. Money was pouring into research at a pace never seen before. The Cold War was, among other things, a war of ideas, and the US government wanted its scientists to be the best in the world.
Computers were getting smaller and faster. The transistor — invented in 1947 — was replacing the fragile vacuum tube. IBM was selling commercial computers to businesses. Programming languages like FORTRAN had just been invented, making computers far easier to use. The future felt genuinely open.
And in the AI community, the Dartmouth Conference of 1956 had just given the field its name and its ambition. Marvin Minsky, John McCarthy, Claude Shannon and others had gathered for a summer workshop and declared that they would make machines that could do anything a human mind could do. The optimism was enormous — and, it would turn out, wildly premature.
What researchers were trying in the late 1950s
After Dartmouth, AI split into two broad camps.
The first camp — which Minsky and McCarthy led — believed in symbolic AI, also called rule-based AI or logic AI. Their approach: write down everything a machine needs to know as logical rules, and let the machine reason from those rules. If the rule book is complete enough and the reasoning engine is powerful enough, you get intelligence.
Early results were exciting. Programs like Logic Theorist (1956) and General Problem Solver (1957) could prove mathematical theorems and solve puzzles. McCarthy wrote a paper called Programs with Common Sense (1958) describing a system that could reason about the world. It felt like intelligence might be just around the corner.
The second camp — smaller, and looked on with some suspicion by the first — believed that the brain was the right model. Not logic — biology. The brain does not run on rules. It is made of neurons, which fire electrochemical signals. It learns by changing the strength of connections between neurons. Maybe that is the right way to build an intelligent machine.
Frank Rosenblatt belonged to this second camp.
The neuroscience that inspired Rosenblatt
In 1943, neurophysiologist Warren McCulloch and logician Walter Pitts had published a paper showing that neurons — modelled as simple on/off switches — could, in principle, compute any logical function. A network of artificial neurons could implement AND, OR, NOT — the basic building blocks of all computation.
This was thrilling. It suggested that the brain was, at its root, a logical machine. And if it was a logical machine, then maybe you could build an artificial version.
But McCulloch and Pitts’ neurons were static — their connections were fixed by hand. They could not learn. Rosenblatt wanted something more: a neural network that would wire itself based on experience.
The key insight came from psychologist Donald Hebb, who in 1949 had proposed a rule for how neurons learn: “neurons that fire together, wire together.” When two neurons are active at the same time, the connection between them gets stronger. When they are rarely active together, the connection weakens. Learning, Hebb argued, is the strengthening and weakening of connections.
Rosenblatt turned Hebb’s idea into mathematics. The result was the Perceptron.
Who was Frank Rosenblatt?
Frank Rosenblatt was a psychologist at the Cornell Aeronautical Laboratory in Buffalo, New York. He was young — 29 when the Perceptron paper appeared — and brilliantly creative, with a flair for the dramatic that made him popular with journalists and controversial among colleagues.
He was not primarily a mathematician or an engineer. He thought of himself as a scientist of the mind, trying to understand how intelligence worked by building it from the ground up. He genuinely believed the Perceptron would, if scaled up, become something like a thinking machine.
The New York Times ran a story in 1958 with the headline: “New Navy Device Learns By Doing.” (The Perceptron was funded by the US Navy.) The Navy was hoping it could be used for image recognition — reading aerial photographs, perhaps identifying enemy aircraft. These applications were never fully realised. But the architecture Rosenblatt invented is, 65 years later, precisely what powers real-time face recognition at border crossings, tumour detection in hospitals, and self-driving cars.
Rosenblatt died in 1971 in a boating accident, at age 43 — before the renaissance of neural networks in the 1980s. He never saw what his idea would become.
Next: The Problem →