What was Frank Rosenblatt's perceptron, and how did it learn?

In the late 1950s, most researchers trying to build thinking machines wrote programs that followed explicit rules. Frank Rosenblatt, a researcher at the Cornell Aeronautical Laboratory in Buffalo, New York, tried something different. He designed a machine made of many simple, neuron-like units whose connections changed with experience. Show it enough examples, and it would learn to tell one pattern from another on its own.

He called it the perceptron. It was one of the first artificial neural networks that could learn, and the ideas behind it sit inside nearly every modern AI system that recognizes images or speech.

When and where it happened

Rosenblatt worked at the Cornell Aeronautical Laboratory with funding from the U.S. Office of Naval Research. The perceptron's learning method is usually dated to 1957, when he first described and simulated it. His best-known paper, "The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain," was received by the journal Psychological Review on April 23, 1958, and appeared in volume 65 later that year.

The first perceptron was not a physical machine. It ran as a simulation on an IBM 704 computer at the laboratory. Rosenblatt then had custom hardware built, the Mark I Perceptron, designed for image recognition.

In 1958 the U.S. Navy held a press conference about the project. Rosenblatt's claims, and the coverage that followed, stirred controversy. The New York Times reported that the Navy expected the device to become a computer able to "walk, talk, see, write, reproduce itself and be conscious of its existence." Many researchers felt that promised far too much.

What question Rosenblatt was asking

The 1958 paper begins not with computers but with the brain. Rosenblatt asked three questions: how a living system senses the physical world, in what form it stores information, and how stored information shapes what it recognizes and does. The first had good answers from physiology. The other two did not.

He sided with the "connectionist" view, the idea that memories are not stored as little pictures somewhere in the brain but in the pattern and strength of connections between neurons. The perceptron was his attempt to show, with mathematics and simulation, that such a system could actually learn.

How the perceptron worked

Rosenblatt's design had three layers of units:

1. Sensory units (S-points): a "retina" of light detectors. In the Mark I, a camera fed a 20 by 20 grid of 400 photocells.
2. Association units (A-units): each one wired, more or less at random, to a handful of sensory points. An A-unit fires when the total signal it receives passes a threshold.
3. Response units (R-units): the outputs, such as "this is a square" or "this is a circle."

Learning happened in the connections leading to the responses. When the machine was rewarded for a correct answer, or corrected for a wrong one, the strengths of the active connections were adjusted. In the Mark I, those strengths were set by potentiometers, the same kind of dial used for volume knobs, turned by small electric motors.

Here is the error-correcting idea in its simplest, textbook form. Suppose a single unit must decide whether a two-pixel image is "bright" (output 1) or "dark" (output 0). It adds up its inputs times their weights, and answers 1 if the total is at least 1:

No one tells the unit what the right weights are. It finds them by being corrected when it gets an answer wrong.

Why it mattered

Rosenblatt's paper reached several conclusions that still sound modern. A network of randomly connected units could learn to link particular responses to particular stimuli. It could generalize, correctly classifying new examples that resembled ones it had learned. And its memory was spread across many connections, so damaging part of the network caused a gradual loss of skill rather than the loss of a single fact. He also admitted a limit: recognizing relationships in space and time seemed beyond what it could abstract.

The excitement did not last. In 1969 Marvin Minsky and Seymour Papert published Perceptrons, a book showing that a single layer of these units can only learn patterns that can be separated by a straight line. It cannot even learn XOR, the rule "one input or the other, but not both." Funding and interest in neural networks fell sharply and recovered only in the 1980s, once researchers showed that networks with several layers are far more capable.

How is the perceptron related to today's deep learning?

A modern deep network is, in a real sense, a very large stack of perceptron-like units. Each unit still multiplies its inputs by weights, adds them up and passes the result through a simple function. Training still means adjusting those weights to reduce errors on examples. The big differences are depth, with many layers instead of one trainable layer, and a training method called backpropagation that can spread corrections back through all those layers.

Scale changed everything. The Mark I had 400 inputs and its weights were knobs turned by motors. Today's image-recognition networks have millions of adjustable weights and train on millions of labeled photos using fast graphics chips. The language model GPT-2, which OpenAI described earlier this month, has 1.5 billion of them.

The line from Rosenblatt to these systems is direct. The question he asked in 1958, whether a machine could learn from experience by changing its connections, is the question deep learning has answered with a much louder yes.