Era 05 / 15 · The Perceptron 1957–1969

The Perceptron

A machine that learns a boundary from examples — then hits a wall.

Beat 1 · Concrete

A line learns to divide

Two groups of dots. A straight line nudges itself until it cleanly splits them.

Two groups of dots — circles and squares — sit in the plane. The perceptron’s learning rule tilts and shifts a single straight line, one mistake at a time, until every circle is on one side and every square on the other: a clean dividing boundary, found from examples alone.

misclassified dot correctly placed clean boundary found ○ group A · □ group B

Beat 2 · Abstract

The unit, stripped bare

Inputs times weights, summed, then a threshold fires. The weights are the boundary.

A perceptron unit driving a decision boundary Two inputs flow through weights w1 and w2 into a summing node, then a threshold step that fires the output. As the weights adjust, the dividing line in the plane below tilts in exact lockstep — the weights and the boundary are the same thing. x₁ x₂ bias w₁ +0.00 w₂ +0.00 Σ step ŷ fire? — same boundary, same weights —

A perceptron unit: each input x is multiplied by a weight w, the products are summed with a bias, and a threshold (step) decides whether the unit “fires”. Adjusting the weights is exactly the same act as tilting the dividing line from Beat 1 — the weights are the boundary.

weights w₁, w₂ (adjusting) boundary they define output fires

Beat 3 · Interactive

Then it hits a wall

Switch the dots to XOR — diagonal pairs. One straight line can never split them.

The same line, but now the dots form an XOR pattern: each group occupies two opposite corners. No single straight line can place all four corners correctly — at least two dots stay misclassified, forever. This is the wall Minsky & Papert proved in 1969, and it set up the first AI winter.

Separable — misclassified: 0

still wrong (the wall) correctly placed line, still trying

Footnotes & further reading
1958
The Mark I Perceptron
Frank Rosenblatt’s machine learned to classify images by adjusting motor-driven weights — hardware that learned a boundary from examples, igniting the first wave of optimism.
1969
Perceptrons
Minsky & Papert proved a single-layer perceptron cannot compute XOR. The result chilled neural-net funding for over a decade.
concept
Linear separability
A perceptron can only separate classes a single straight line (hyperplane) can divide. XOR isn’t linearly separable — hence the wall.