Perceptrons: An Introduction to Computational Geometry
One straight line cannot separate XOR — and that geometric wall marks the limit of a single-layer perceptron.
Why could the first learning machine tell an A from a B, yet stumble on a puzzle a child finds easy — is this dot the same as that one, or different?
The big idea
A perceptron — the first machine that learned from examples (see Rosenblatt, 1958) — draws a single straight line to sort things into two groups: yes on one side, no on the other. That works for any question a line can answer. Minsky and Papert asked the sharp follow-up: which questions can no line answer?
Their famous case is exclusive-or, or XOR: "one or the other, but not both." Put the four yes/no combinations on the corners of a square, and the two "yes" answers sit on opposite corners, with the two "no" answers on the other pair. Try as you might, no single straight line can put both yeses on one side and both nos on the other. So a one-line machine simply cannot learn XOR — not because it was trained badly, but because the shape of the problem forbids it.
How it came about
By the late 1960s, perceptrons had been sold with extraordinary fanfare — newspapers spoke of machines that would soon walk, talk, and be conscious. Marvin Minsky and Seymour Papert, two founders of artificial intelligence at MIT, set out to replace the hype with proof. Their 1969 book used careful geometry to map exactly what a single-layer perceptron could and could not do, and the verdict was sobering: a whole class of natural problems, XOR among them, lay permanently out of reach.
The book was right, and its impact was enormous — arguably too enormous. Read as a judgment on neural networks in general, it helped dry up funding for the field for more than a decade, the period now called the first "AI winter." Frank Rosenblatt, the perceptron's inventor and Minsky's old schoolmate and rival, died in a boating accident in 1971, before the idea was revived.
Why it mattered
It forced a distinction that the field had blurred: the difference between a machine that learns slowly and a machine that cannot represent the answer at all. No amount of training fixes a missing capability. That lesson — know the limits of your model's shape before you blame the training — is now bedrock in machine learning.
It also pointed, by contrast, at the cure. If one line is not enough, use more: stack simple units into layers, and let later layers combine the lines the earlier ones draw. The wall Minsky and Papert built was the exact shape of the door that backpropagation (Rumelhart, 1986) would later open.
A way to picture it
Imagine sorting a pile of socks by laying a single straight stick on the floor: plain socks to the left of the stick, striped to the right. If the plain and striped socks happen to be mixed at opposite corners of the pile, no straight stick can separate them — wherever you lay it, a wrong sock ends up on each side. You don't need a better stick; you need a second stick, and a rule that combines them. That second stick is a hidden layer.
Where it sits
This is the middle chapter of the Library's neural-network story. It opens with Rosenblatt's perceptron (1958), the first machine to learn from examples; Minsky and Papert (1969) draw its hard boundary; and Rumelhart, Hinton, and Williams (1986) cross that boundary with backpropagation, training the hidden layers that make XOR — and far harder things — easy. Everything after, from AlexNet (2012) to the Transformer (2017), lives on the far side of the wall this book made precise.
One popular version is that the publication of our book so discouraged research on learning in network machines that a promising line of research was interrupted. Our version is that progress had already come to a virtual halt because of the lack of adequate basic theories.