The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain

Frank Rosenblatt's paper "The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain" appeared in Psychological Review in 1958, volume 65, number 6, pages 386 to 408.^[1] The work was carried out at the Cornell Aeronautical Laboratory under the sponsorship of the U.S. Office of Naval Research.^[2]

The problem

Rosenblatt began with three questions: how information about the physical world is sensed, in what form it is stored or remembered, and how stored information influences recognition and behaviour.^[2] His answer took the "connectionist" side. Memory, he argued, is not stored as coded images of what was seen. It lies in new connections or pathways between cells. He cited Donald Hebb's work as a major source of inspiration.^[2]

This was a different picture from the one in which memories are stored as copies of what was seen. The difference matters for learning: if knowledge lives in connections, changing those connections is a way of changing what the system does.

The method: the photoperceptron

The model Rosenblatt called the "photoperceptron" has three kinds of parts. A retina of sensory units (S-points) receives the stimulus. One or more layers of association cells (A-units) are connected to the retina. Response units (R-units) produce the output. Learning changes the values of the A-units connected to each response.^[2]

A simple way to picture training is to show the machine a stimulus, record its answer and compare that answer with the correct one. When it is wrong, the connections that helped produce the wrong answer are adjusted, so the next attempt is more likely to be right. Repeating this many times is the trial-and-error training the paper analyses.

The results

The paper analyses mathematically how such a system learns to classify stimuli through trial-and-error training, and it predicts learning curves from the system's parameters. Rosenblatt places his model alongside earlier brain models, citing McCulloch and Pitts (1943) among them.^[2]

The model was also demonstrated, not only analysed. In July 1958 the Office of Naval Research publicly showed the perceptron simulated on an IBM 704 computer. After 50 trials, the machine learned to tell punch cards marked on the left from cards marked on the right.^[3]

Why it mattered

The perceptron showed that a machine could improve its classifications from examples, rather than relying only on rules that people wrote out in advance. That shift from writing every rule by hand to training a system with examples is central to the neural networks that followed. The paper supports the idea with mathematics as well as with a demonstration.^[2]

What followed

The perceptron became the starting point for trainable neural networks. Its limitations were later analysed in Minsky and Papert's 1969 book Perceptrons.^[3]

The name "perceptron" later came to stand for a family of trainable units and networks, and the single-layer machine described in this paper is its simplest member.