In the Early Days of Machine Learning, Massive Computers Said George Harrison Was a Woman. A.I. Has Come a Long Way

A Cornell professor designed a room-size network of sensors that represented a single neuron. He claimed it would grow wiser as it gained experience, and it has never stopped

Jennie Rothenberg Gritz

Illustration by Carmen Martínez

Summer 2026

Share

Copy linkEmailSMSFacebookXRedditLinkedInBlueskyPrint

Add as preferred source

“Electronic ‘Brain’ Teaches Itself.” That was how the New York Times defined the Mark I Perceptron in a headline on July 13, 1958. Developed at Cornell University by psychologist Frank Rosenblatt (who turned 30 two days before the article appeared), the Perceptron was a room-size grid of 400 light--registering sensors designed to perceive and sort images. The idea that a computer could “see” was revolutionary enough, but Rosenblatt also told the Times that the machine was designed to “grow wiser as it gains experience.” 

The Perceptron was the first step toward the kind of artificial intelligence that’s ubiquitous today, where machines “learn” from experience in a way that mimics human brains. Like so many other advances of the 1950s, its early history now reads like a mid-century science fiction novel. When the Perceptron debuted, “people heard, ‘You can build a brain. It can do everything that humans can do,’” says Kilian Weinberger, a computer science professor at Cornell who teaches classes on machine learning. But scientists had a rudimentary understanding of the human brain. They weren’t able to study healthy, living gray matter using technologies like magnetic resonance imaging (MRI). Instead, they relied on indirect approaches such as dissecting corpses or observing the effects of brain lesions on patients’ behavior.

A theory began to emerge: New experiences strengthened the connections between brain cells. Once these networks formed, the brain weighed new information differently. For instance, after a toddler was told over and over that a round object was a ball, the child would quickly and intuitively recognize other round objects as balls, too.

Rosenblatt designed the Perceptron to mimic this process. His machine took in outside information, processed it through circuits and transmitted the resulting answer. Researchers trained the machine by using yes/no feedback that updated how it weighed incoming information. Black-and-white footage from the mid-1960s shows this type of training in action. A researcher sits in front of the giant machine, “teaching” it how to distinguish men from women. He projects slides of faces and uses two simple switches labeled “man / woman” and “wrong / right.” When the machine identifies George Harrison of the Beatles as a woman, based on his shaggy hairstyle, the researcher gives the feedback “wrong.” Through trial and error, the machine will learn that men can have long hair, too.

Everything about this example seems quaint today, but the real limitation was that the Perceptron mimicked a single neuron. “One neuron was actually very impressive, given the hardware they had at the time,” Weinberger says. “It took a whole room full of equipment and a lot of work.” Still, the public eventually realized that the much-touted “electronic brain” was less powerful than the brain of a mouse. 

Rosenblatt died in a boating accident in 1971 at the age of 43. When David Tank began his PhD program in physics at Cornell later that decade, the Perceptron and its creator were already the stuff of legend. Tank remembers going to a pig roast out in the country, where the host took him to a barn and showed him an experimental Perceptron machine built in Rosenblatt’s lab: “It had all of these tubes and wires, and it was just gathering dust.” Around the same time, Tank read Rosenblatt’s 1962 book, Principles of Neurodynamics. “It was basically the first textbook on artificial neural networks, and it’s an inspirational book in many ways, an absolute classic,” says Tank, who became a leading A.I. innovator at Bell Labs and co-founded the Princeton Neuroscience Institute.

What happened between the Beatles haircut lesson and the age of modern A.I.? For years, it seemed as though other forms of machine intelligence would win out. Rule-based logic, for instance, used individually coded “if-then” instructions to build an elaborate flow chart. The machine could obey those commands to find a solution, but it couldn’t learn. Neural networks briefly entered the chat again in the late 1980s, when researchers collaborated with the U.S. Postal Service to develop ZIP code recognition. A team at Bell Labs, led by Yann LeCun, successfully trained machines to recognize different types of handwriting. Instead of foreseeing and specifying every possible way a person could close the loop of an 8 or angle the line of a 7, LeCun and his colleagues trained their system on many examples until it got the hang of human handwriting. 

Still, neural networks weren’t able to advance much until computers had more storage and bandwidth. A crucial development came through the computer games industry in the form of graphics processing units, or GPUs—specialized circuits that generate images and videos. These networks quickly and accurately performed thousands of calculations at once, in a way that finally brought truly useful machine learning within reach. 

The other game changer was, of course, the internet. Forget about a lone scientist inputting photos of shaggy-haired musicians. A.I. could now draw on sources ranging from ancient Sanskrit scriptures to scientific papers to massive libraries of images and videos to endless snarky social media threads. The new form of A.I. could learn an unfathomably large amount about art, science, literature and human nature and weight its results accordingly. 

When Rosenblatt introduced his Perceptron in 1958, he wrote in a research paper that he wanted “to understand the capability of higher organisms for perceptual recognition, generalization, recall and thinking.” His original Mark I machine, now in the Smithsonian’s collections, is ultimately a testament to the supercomputer inside the human skull, with its relentless urge to expand its own powers—maybe for worse and maybe for better. 

Jennie Rothenberg Gritz | Read More

Jennie Rothenberg Gritz is a senior editor at Smithsonian magazine. She was previously a senior editor at the Atlantic.

Carmen Martínez | | Read More

Carmen Martínezis a Spanish editorial illustrator known for vibrant, symbolic imagery exploring identity and human connection.