The History of AI

Perceptrons

Era 2 · AI is born (1943–1979) › Neural networks and learning

Lineage
Neurons and networks
Kind
Publication

Who

MIT

What happened

Marvin Minsky and Seymour Papert of MIT publish Perceptrons: An Introduction to Computational Geometry with the MIT Press. An expanded edition with a new prologue and epilogue appears in 1988.

Why it mattered

They prove rigorously what a single-layer perceptron whose units see only part of the input cannot compute, such as parity or whether a figure is connected. The book is blamed for the decline in connectionist funding, a reading that is disputed.

The full story

In the late 1960s the perceptron was competing with symbolic artificial intelligence for attention and funding. Minsky, who had built SNARC, and Papert, a mathematician who had worked with Piaget, decided to study it as a mathematical object: which functions it can recognize and how many connections it needs.

Their results are precise. A perceptron whose intermediate units each look at only a limited number of points in the image cannot decide whether the number of active points is even or odd, parity, which in its smallest form is the XOR function. Nor can it decide whether a figure is connected unless some unit sees a part of the image that grows with the figure's size. The book also examines how much memory and how much learning time certain problems would require.

The proofs concern architectures with a single adjustable layer. The book acknowledged that multilayer networks might be more powerful but doubted they could be trained. Methods for doing so already existed, such as Ivakhnenko's in Kyiv or Amari's in Japan, but they had been published far from the community that followed Rosenblatt, where the problem was considered open.

For years the book was blamed for sinking neural network research in the 1970s. The authors denied it, and critics such as Jordan Pollack argue that a narrow result was read as a general condemnation. The 1988 edition responded to the revival of neural networks without new results.

Two plots with the four cases of two inputs, a and b. In the AND plot only the point (1, 1) is filled, and a line separates it from the other three. In the exclusive OR plot (0, 1) and (1, 0) are filled, on a diagonal, and two trial lines fail to separate them from the empty points.

A straight line separates the cases of AND, but none separates those of exclusive OR.

Under the hood

A single-layer perceptron draws a straight boundary: it outputs 1 on one side of w₁a + w₂b + c = 0 and 0 on the other. For AND, the line a + b = 1.5 will do.

For exclusive OR, these would all have to hold at once: c ≤ 0 (it outputs 0 at (0, 0)), w₁ + c > 0 and w₂ + c > 0 (it outputs 1 at (1, 0) and (0, 1)), and w₁ + w₂ + c ≤ 0 (it outputs 0 at (1, 1)). Adding the two middle conditions gives w₁ + w₂ + c > −c ≥ 0, which contradicts the last one. No weights will work.

With one intermediate unit that computes "a AND b," the problem is solved: the output is 1 if a + b − 2·(a AND b) > 0.5. The book treats the general case, the parity of a whole retina, and shows that it requires units that see the entire image.

Common misconceptions

  • What people often say: Perceptrons proved that neural networks cannot compute functions like exclusive OR.

    In fact: Its proofs concern perceptrons with a single adjustable layer. A network with a hidden layer solves exclusive OR; what Rosenblatt's community lacked was a known method for training it, and Ivakhnenko's and Amari's methods already existed.

Builds on

  • The perceptron (1957)

    Documented direct influence

    The book analyzes the perceptron mathematically and is named after it

Threads that pass through here

Sources

How to cite

“Perceptrons.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/minsky-papert-perceptrons

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline