Perceptron
In short
A perceptron is the simplest artificial neural network: one unit that takes several inputs, weighs them and gives a yes-or-no answer. Frank Rosenblatt introduced it in 1958, and it learns by correcting itself when it gets an answer wrong.
- Also called
- Rosenblatt perceptron, Mark I Perceptron, single-layer perceptron
- Topic
- How AI models work
- Level
- Beginner
- First described
- Frank Rosenblatt of the Cornell Aeronautical Laboratory, in "The perceptron: a probabilistic model for information storage and organization in the brain", Psychological Review, 19581.
The perceptron was set out by Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory in Buffalo, New York, in a 1958 paper in Psychological Review1. The work was funded by the US Office of Naval Research1. Rosenblatt's aim was to model how a brain might store information and use it to recognise things1.
It sorts an input into one of two categories, such as left or right, or dog or cat. It makes a prediction and, if the prediction is wrong, adjusts itself so the next guess is better2. When it gets an answer right, nothing changes: the perceptron only updates on mistakes3.
In July 1958 the US Office of Naval Research showed it off on an IBM 704 computer. After 50 trials the machine had taught itself to tell punch cards marked on the left from cards marked on the right2. The press was excited. A New York Times headline read "New Navy device learns by doing"2.
The catch was that Rosenblatt's perceptron had a single layer, and he did not know how to train networks with several layers2. In their 1969 book Perceptrons, MIT's Marvin Minsky and Seymour Papert set out its limits, including that it cannot handle the "exclusive or" (XOR) logic problem. Their critique made neural networks unfashionable for years4.
Today's deep learning systems rest on the same basic idea, scaled up to many layers2. The original perceptron still ships as a simple classifier in the scikit-learn machine learning library3.
Example
Picture a perceptron deciding whether an email is spam. Its inputs might be "mentions a prize" and "sender is in your contacts". Each input carries a weight; if the weighted total passes a threshold, the answer is spam. If it wrongly flags a message from your bank, it shifts the weights slightly and tries again on the next email.
Why it matters
It is the starting point for neural networks, the family of methods behind today's chatbots. Its story is also an early case of AI hype followed by a long slump.
Questions people ask
Who invented the perceptron?
Frank Rosenblatt, a psychologist working at the Cornell Aeronautical Laboratory, described it in a 1958 Psychological Review paper1. The US Office of Naval Research demonstrated it on an IBM 704 computer in July 19582.
Why can't a perceptron solve XOR?
XOR, or "exclusive or", means one input or the other is true but not both. Minsky and Papert's 1969 book Perceptrons showed that Rosenblatt's design had limited ability with such logic problems, XOR being the best-known case4. Rosenblatt's machine had a single layer, and at the time he did not know how to train networks with several layers2.
Is a perceptron the same as a brain neuron?
No. It was inspired by the way neurons work together in the brain2, but it is a simple mathematical rule for sorting inputs into two groups, not a model of real biology.
Related terms
Sources
- Rosenblatt (1958), The perceptron: a probabilistic model for information storage and organization in the brain, Psychological Review 65(6)
- Cornell Chronicle (25 September 2019), Professor's perceptron paved the way for AI, 60 years too soon
- scikit-learn documentation, Linear models: Perceptron
- Lehigh University Libraries exhibit, Minsky and Papert (1969), Perceptrons: an Introduction to Computational Geometry