In the early days of machine learning, giant computers said George Harrison was a woman. AI has made great progress

“The electronic ‘brain’ learns by itself. “That’s why, new york times defined the Mark I Perceptron under the heading July 13, 1958. Developed at Cornell University by psychologist Frank Rosenblatt (who turned 30 two days before the article was published), the Perceptron was a sensor designed to recognize and classify images in a room-sized grid of 400 lights. The idea that computers could “see” was revolutionary enough, but Rosenblatt also said: times The machine is designed to “get smarter as it gains experience.”

Perceptrons were the first step towards the kind of artificial intelligence that is ubiquitous today, where machines “learn” from experience in a way that mimics the human brain. Like many other advances of the 1950s, its early history now reads like a mid-century science fiction novel. When the perceptron debuted, “people heard, ‘We can build a brain. We can do everything a human can do,'” says Kilian Weinberger, a computer science professor at Cornell University who teaches a class on machine learning. However, scientists had a rudimentary understanding of the human brain. They were unable to study healthy, living gray matter using techniques such as magnetic resonance imaging (MRI). Instead, they relied on indirect approaches such as dissecting cadavers and observing the effects of brain lesions on patients’ behavior.

Theories began to emerge that new experiences strengthen the connections between brain cells. Once these networks were formed, the brain began to evaluate new information differently. For example, if a young child is taught repeatedly that a round object is a ball, the child will soon intuitively recognize other round objects as balls as well.

Rosenblatt designed the perceptron to mimic this process. His machine took in external information, processed it through circuitry, and transmitted the resulting answer. The researchers trained the machine using “yes/no” feedback that updated how it weighted incoming information. Black and white footage from the mid-1960s shows this type of training in action. Researchers sit in front of giant machines and “teach” them how to distinguish between men and women. He projects a slide of his face and uses two simple switches labeled “Male/Female” and “Wrong/Correct.” When a machine identifies Beatle George Harrison as a woman based on his shaggy hairstyle, researchers give feedback that it’s “wrong.” Through trial and error, the machine learns that men can grow their hair long too.

All of this example seems strange today, but the real limitation was that the perceptron imitated a single neuron. “Considering the hardware at the time, certain neurons were actually quite impressive,” 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 capable than a rat’s brain.

Rosenblatt died in a boating accident in 1971 at the age of 43. When David Tank began his doctoral program in physics at Cornell University later that decade, the perceptron and its creator were already the stuff of legend. Mr. Tank remembers going to a pig roast in the countryside. There, his master took him to a barn and showed him an experimental perceptron machine built in Rosenblatt’s lab. “It had all these tubes and wires and it was just gathering dust.” Around the same time, Tank read Rosenblatt’s 1962 book. Principles of neurodynamics. “This is basically the first textbook on artificial neural networks, and in many ways it’s an inspirational book and an absolute classic,” says Tank, who went on to become a leading AI innovator at Bell Labs and a co-founder of the Princeton Neuroscience Institute.

What happened between the Beatles’ haircut lessons and the modern age of AI? For years, it seemed like other forms of machine intelligence would win. For example, rule-based logic used individually coded “if-then” instructions to construct elaborate flowcharts. The machine can follow these commands to find a solution, but it cannot learn. Neural networks briefly reappeared in chat in the late 1980s, when researchers worked with the U.S. Postal Service to develop zip code recognition. A Bell Labs team led by Yann LeCun has successfully trained a machine to recognize different types of handwriting. Rather than predicting and specifying all the ways a person could close a ring of 8 or angle a line of 7, LeCun and his colleagues trained the system on many examples until it mastered the tricks of human handwriting.

Still, neural networks couldn’t make much progress until computers had more storage and bandwidth. An important development in the computer gaming industry came in the form of the graphics processing unit (GPU), a specialized circuit that generates images and video. These networks can now quickly and accurately perform thousands of calculations at once, ultimately enabling truly useful machine learning.

Another game changer was, of course, the internet. Forget about the lone scientist typing pictures of unkempt-haired musicians. AI can now tap into a variety of sources, from ancient Sanskrit scriptures to scientific papers, vast libraries of images and videos, and endless vile social media threads. New forms of AI have the potential to learn immeasurable amounts about art, science, literature, and humanity, and weight outcomes accordingly.

When Rosenblatt introduced the perceptron in 1958, he wrote in a research paper that he wanted to “understand the perceptual recognition, generalization, recall, and thinking capacities of higher organisms.” His original Mark I machine, now housed in the Smithsonian Institution, is evidence of a supercomputer that ultimately resides inside the human skull, with a relentless urge to expand its capabilities, perhaps for the worse, and perhaps for the better.