1958: The Perceptron makes headlines

On 7 July 1958, the U.S. Navy held a press conference in Washington, D.C. to unveil an unusual new machine: the Perceptron. Built by psychologist Frank Rosenblatt at the Cornell Aeronautical Laboratory in Buffalo, New York, the Perceptron was an early artificial neural network designed to recognize simple visual patterns, such as distinguishing shapes marked on cards. The demonstration was covered the very next day by the New York Times in an article titled "New Navy Device Learns By Doing," which described a machine the Navy expected would eventually be able to walk, talk, see, write, reproduce itself, and be conscious of its own existence. Those predictions were, to put it mildly, premature.

The Perceptron itself was a relatively modest piece of hardware: a room-sized array of photocells, potentiometers, and motors simulating a single layer of adjustable connections, loosely inspired by neurons in the brain. Rosenblatt's real contribution was theoretical. He showed that a simple network could learn from examples by adjusting the strength of its connections, a process that anticipated the training methods used in modern machine learning. Crucially, this was one of the first public demonstrations of a machine that improved its own performance through experience rather than following a fixed set of rules.

The excitement did not last. In 1969, Marvin Minsky and Seymour Papert published "Perceptrons," a rigorous mathematical critique showing the sharp limitations of single-layer networks, notably their inability to solve problems like XOR. This contributed to a decline in funding and enthusiasm for neural network research, part of what later became known as the first "AI winter."

The story, however, was far from over. Decades later, researchers found ways around these limitations by stacking multiple layers together and developing better training algorithms, most notably backpropagation. This layered approach eventually became deep learning, the foundation of today's image recognition, language models, and countless other AI systems.

Today's neural networks are vastly more complex than Rosenblatt's original machine, but the core idea, that a system can learn by adjusting internal connections based on data, traces directly back to that press conference on 7 July 1958. It stands as an early reminder that in AI, both the hype and the eventual payoff can take much longer to arrive than anyone expects.