What is a perceptron? A 1958 artificial neuron that adjusts its own weights when it gets an example wrong

A 1958 artificial neuron that adjusts its own weights when it gets an example wrong.

The perceptron is an artificial neuron introduced in 1958 by American psychologist Frank Rosenblatt. It multiplies each incoming signal by a weight, adds them up, and outputs 1 if the total clears a threshold, 0 if not. The new part: whenever it got an example wrong, it adjusted its own weights a little.

A single-layer perceptron has clear limits. In 1969 Marvin Minsky and Seymour Papert proved mathematically that it can't solve problems such as XOR, which can't be split by one straight line. Neural network research lost momentum for years, until stacking layers and training them with backpropagation caught on.

Think of it like this

Think of a scoring sheet that adds up a few scores and says "pass" if the total clears a cutoff. A perceptron is a scoring sheet that changes how much each score counts whenever it makes a wrong call. But one cutoff line can only capture simple rules.

For example

XOR, "output 1 only when exactly one of two switches is on", can't be solved by a single-layer perceptron.

Common mix-ups

Mix-upThe perceptron was the first AI that could learn anything.
ActuallyRosenblatt's 1958 paper appeared not in a computing journal but in the psychology journal Psychological Review, and was closer to a model of how the brain learns. The name "artificial intelligence" was already in use before it, and with only one learning layer it could handle only certain kinds of problems.
Mix-upOnce the perceptron's limits were shown, neural network research ended.
ActuallyWhat Minsky and Papert pinned down mathematically was the limit of a single layer. Stacking layers could represent XOR, but there was no good way yet to train several layers, so the research shrank for a decade or so rather than disappearing.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: Why can't a single cut-off line capture a rule like XOR?
Activity: On grid paper, plot the four states of two switches (off-off, on-off, off-on, on-on) and mark the points whose result is 1 differently from the 0s. In groups, use a ruler to test whether AND and OR can be split by one straight line — and then XOR. Try two lines and you'll see what "stacking layers" means.