SAEDNEWS: In 1958, Cornell psychologist Frank Rosenblatt introduced the Mark I Perceptron, a room-sized machine designed to learn from experience. Its limitations were enormous, but the ideas behind it helped lay the groundwork for today’s neural networks and A.I.
According to SaedNews: In July 1958, the New York Times announced an unusual new machine under the headline “Electronic ‘Brain’ Teaches Itself.” The device was the Mark I Perceptron, created at Cornell University by psychologist Frank Rosenblatt, who had turned 30 just two days earlier. Filling an entire room, the machine contained a grid of 400 sensors capable of registering light and was built to recognize and classify images. Rosenblatt told the Times that the machine was intended to “grow wiser as it gains experience.”
The Perceptron represented an early attempt to reproduce the way the human brain learns. At the time, however, scientists had only a limited understanding of the brain itself. They could not examine healthy living gray matter with technologies such as MRI and instead relied on methods including studying corpses and observing how brain injuries changed behavior.
One emerging theory held that experience strengthens connections among brain cells and changes the way the brain evaluates new information. A child repeatedly taught that a round object is a ball, for example, eventually learns to recognize other round objects in the same way.
Rosenblatt attempted to reproduce that process electronically. The Perceptron received information, processed it through circuits and produced an answer. Researchers then supplied yes-or-no feedback, allowing the machine to adjust the weight it assigned to incoming information.
A surviving film from the mid-1960s shows the machine being trained to distinguish men from women. A researcher projected photographs of faces and operated switches marked “man / woman” and “wrong / right.” When the system identified Beatles guitarist George Harrison as a woman, apparently influenced by his shaggy hairstyle, the researcher marked the answer “wrong.” Repeated trials were meant to teach the machine that men could also have long hair.
The demonstration seems almost charmingly primitive now, but the deeper problem was the Perceptron’s limited scale. It essentially modeled a single neuron. “One neuron was actually very impressive, given the hardware they had at the time,” says Cornell computer science professor Kilian Weinberger. Yet accomplishing that required a room full of equipment and considerable effort. The public eventually recognized that the celebrated “electronic brain” was less capable than a mouse’s brain.
Rosenblatt died in a boating accident in 1971, aged 43. Later that decade, David Tank arrived at Cornell for a PhD in physics and found that the Perceptron and its inventor had already become legendary. At a country pig roast, Tank was taken to a barn where an experimental Perceptron built in Rosenblatt’s laboratory sat unused. He recalled seeing “all of these tubes and wires” gathering dust.
Tank also read Rosenblatt’s 1962 book, Principles of Neurodynamics. He later described it as essentially the first textbook on artificial neural networks and “an inspirational book in many ways, an absolute classic.” Tank went on to become a major A.I. innovator at Bell Labs and co-founder of the Princeton Neuroscience Institute.
For years after the Perceptron, other approaches to machine intelligence appeared more promising. Rule-based systems relied on individually programmed “if-then” instructions. They could follow elaborate commands to reach an answer, but they could not learn from experience.
Neural networks returned to prominence in the late 1980s. Researchers worked with the U.S. Postal Service on recognizing ZIP codes, while a Bell Labs team led by Yann LeCun trained machines to identify different forms of handwriting. Instead of programming every possible way someone might draw an 8 or complete a 7, LeCun’s group exposed the system to numerous examples until it learned the patterns.
Yet neural networks remained constrained by limited computer storage and bandwidth. A major breakthrough came from the video-game industry with graphics processing units, or GPUs, specialized circuits originally designed to produce images and video. Their ability to perform thousands of calculations simultaneously made practical, powerful machine learning increasingly feasible.
The internet supplied another transformative ingredient: enormous quantities of data. Rather than relying on one researcher to feed photographs into a machine, A.I. systems could draw information from sources as varied as ancient Sanskrit texts, scientific publications, vast collections of images and videos, and countless social-media discussions. These systems could absorb immense amounts of information about art, science, literature and human behavior and adjust their results accordingly.
When Rosenblatt unveiled the Perceptron in 1958, he wrote that his goal was to understand the ability of higher organisms to recognize, generalize, recall and think. His original Mark I machine, now held by the Smithsonian, remains a physical reminder of those early ambitions—and of the extraordinary learning system inside the human brain.
