Rosenblatt's Perceptron: How the History of Artificial Intelligence Began

In 1958, the American psychologist Frank Rosenblatt unveiled the perceptron, a device many consider the first practically working neural network. It was a decisive step that opened the way to modern artificial intelligence.

What is a perceptron

The perceptron was an electronic machine that imitated the workings of a biological neuron in the brain. Visual information (for example, an image of an object or a letter) was fed in and converted into electrical signals through an array of photocells. The device contained electronic components and a system of weights that could be changed mechanically. Each weight determined how strongly one input signal influenced the machine's output decision.

The main innovation was that the perceptron could learn. The learning algorithm was simple: if the machine made a mistake, the weights were adjusted. Over time, the perceptron grew more and more accurate at recognizing patterns. It resembled the way the human brain learns through repetition and error correction.

Rosenblatt's story and the birth of neural networks

Frank Rosenblatt (1928–1971) was an American psychologist and computer scientist who worked at the Cornell Aeronautical Laboratory. His project received backing from the U.S. Navy, which saw huge potential in automating the analysis of intelligence information and imagery.

The first perceptron, called Mark I, was a major feat of engineering. It was shown to the public in 1958 at a press conference attended by journalists, scientists and senior Navy officers. Rosenblatt confidently declared that the perceptron marked the beginning of an entirely new era of machine intelligence that would soon put an end to routine intellectual work.

Physically, the device looked like a wall of electronics the size of a refrigerator. Inside were 400 photocells for perceiving the image, a complex system of electronic circuits and mechanical switches for adjusting the weights.

Capabilities and limitations

The perceptron could solve tasks that were impressive for the time, though by today's standards they look elementary:

• Recognize simple geometric shapes (squares, triangles, circles) with accuracy above 90%
• Tell apart letters of the Latin alphabet with enough accuracy for practical use
• Learn from a small set of examples without explicit programming of each recognition rule
• Adapt to minor variations in input data (rotation, shifting, scaling)
• Work in real time, an outstanding technical achievement for 1950s electronics

But in the late 1960s, the mathematician Marvin Minsky and his colleague Seymour Papert published a critical work proving fundamental limitations of the perceptron. They showed that a single-layer perceptron cannot solve even simple logical tasks such as the XOR function. The criticism led to a sharp drop in interest in, and funding for, neural network research. What later came to be called the "AI winter" had begun.

What this means

Rosenblatt's perceptron was not merely an interesting engineering find; it was proof that machines can learn in practice. Before the perceptron, all computers were simply calculators executing programmed instructions. The perceptron showed that an electronic device can change its behavior based on experience.

Today, more than 60 years later, modern neural networks are direct descendants of Rosenblatt's perceptron. Yes, they contain billions of parameters instead of dozens, run on graphics processors instead of electromechanical switches, and use sophisticated backpropagation algorithms instead of the simple Hebbian rule. But the basic idea remains the same: a system of weights that adapts to data and learns to solve tasks from examples.