Can computers learn by themselves, just by looking at thousands of examples?
It took me a while to understand what this means. Computers are programmed with lots of rules and algorithms, you need to tell the machine exactly what to do. But when I first started to learn about neural networks, I was blown away by the beauty and simplicity of this idea. And I was even more surprised to learn it all started as far back as 1958 … with a machine called the Perceptron.
Think back to the 1950s. The Cold War was at its height. And the US military had a massive problem. High altitude spy planes (such as the U2) and early reconnaissance satellites captured millions of aerial photographs of Soviet territory. All these images had to be analysed, and the military were literally drowning in imagery which they could not process fast enough. The work of looking at all these photographs and determining if there was any military value in them (e.g. photos showing enemy tanks, aircraft, and geographical targets hidden in complex backgrounds) completely relied on human experts. The military specifically wanted to sort objects into different categories based strictly on visual inputs, what is nowadays called a classification problem. This could involve applications such as distinguishing the silhouettes of Allied naval ships from Soviet naval warships (Friend vs. Foe) or spotting the specific outlines of enemy tanks, transport vehicles, or aircraft sitting on runways. Or identifying man-made infrastructure such as disguised missile silos, submarine docks, or radar installations, hidden against complex, natural terrain backgrounds.
The military could not use standard digital computers of the 1950s for this task, even though they existed and were already performing many other tasks. Traditional computers required explicit, hand-coded rules to function. If a programmer wrote a code to find a tank based on a side-profile view, the computer would fail completely if a picture of the tank was taken from the front, or from the other side. It failed if the tank was partially hidden by a tree, or shadowed by clouds. Pattern recognition did not work very well with heuristic rules. A new approach was needed.
From the early days of building AI systems there were two possible paths to achieving artificial intelligence: figure out how "intelligence" works by discovering the rules which make it work ... which would turn out to be expert systems. It was the most promising approach initially and delivered the most progress. But expert systems hit a wall (see my 2nd post). These systems could not learn anything new. So what was the second avenue? We already have many examples of "intelligence" existing out in the wild. Nature has found a way to make living beings (us included, believe it or not) intelligent, so how does Nature do it? What if we could replicate the wet biology of the brain inside a machine? This is what Frank Rosenblatt set out to do.
Rosenblatt (1928–1971) was a Cornell-trained psychologist and polymath, who had also trained in statistics and neurobiology. He was therefore well-placed to envision how a neuron in the brain receives electrical signals from other neurons connected to it, and then transmit these signals on to the next neurons it is connected to. The strength of the signals exchanged between neurons are related to the importance of the connection: learning develops and grows by strengthening connections which are useful, and weakening connections which are not important. Rosenblatt managed to take existing mathematical models of neurons and collections of such neurons (called neural networks) and build a functioning model of such a network in reality.
With the help of US Navy funding and working at the Cornell Aeronautical Laboratory in Buffalo (New York state, USA) he managed to build an electrical "brain" in 1958, which he called the Perceptron. The brain in question was a single-layer trainable neural network that adjusted its connection weights based on errors, and it could learn to distinguish triangles from squares, the letter E from the letter F, cards marked left from cards marked right. I was watching pictures of this machine, and it looks like an invention straight out of a sci-fi movie, a bunch of wires and motors and switchboards, and scientists standing around the machine with white shirts and black ties. Even today it still looks futuristic! The most amazing thing is that the machine actually worked … nobody had programmed the machine with the rules to distinguish between a letter "E" and "F", but show it enough examples, and it started to figure it out on its own. The brain had learned.
The initial results were very promising, and a real hype broke out. Partly as a result of pressure by the military to show results, press articles called it the beginning of a new era. The New York Times, on July 8, 1958, reported the Perceptron as "the embryo of an electronic computer that the Navy expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence". Rosenblatt told reporters that Perceptrons might eventually be fired to other planets as mechanical explorers. The headline read: "NEW NAVY DEVICE LEARNS BY DOING." The New Yorker gushed in turn.
Quite heavy claims for a machine which could barely separate simple, high-contrast geometric shapes like white triangles from squares on a black background. As often happens in AI's exciting history, real progress was oversold, and sky-high expectations could not be realized. Soon the hype died down. The Perceptron was not able to learn the abstract concept of a "tank" on its own, allowing it to spot them under varying, real-world field conditions. And the single layer neural network which the Perceptron was built on never could recognize more complex patterns, a fact discovered by Minsky and Papert at MIT in 1969. It was the end of the Perceptron. Rosenblatt himself never recovered, and he sadly passed away in 1971 during a sailing accident.
And yet ... his legacy lives on. LLMs are based on Perceptrons, but it would need many additional developments to make them genuinely useful. For a starter it was recognized that a single layer neural network would never be able to recognize complex patterns, a multi-layer Perceptron (MLP) was needed. But in the 1960s and 1970s there was no algorithm available to calculate the parameters of such an MLP and deal with the complexity of the multiple layers. It would take until 1986 to solve that problem. The second barrier was the limitations on the hardware available, computers were just not powerful enough to perform the many calculations required to recognize complex patterns. And the third obstacle was the severe limitation of data ... any neural network can only learn by feeding it with many examples of the patterns you want it to recognize, and the datasets required simply didn't exist.
The Perceptron failed in the short term, but it planted a seed that took decades to bloom. What strikes me most is not the hype or the crash, but the stubbornness of the idea itself: that a machine could learn by doing, not by being told. Rosenblatt didn't live to see it vindicated. But the next time you get a useful answer from your favourite AI tool, somewhere in that response is a faint echo of that wonderful machine built in a Buffalo lab, back in 1958.
The three obstacles that doomed the Perceptron, the missing algorithm, the weak hardware, and the absence of data, each fell in time. In the next post I'll pick up where this one leaves off. If this story surprised you as much as it surprised me, drop me a note, I'd genuinely love to hear your reaction.