The Perceptron and Its First Hype Cycle

In 1958, psychologist Frank Rosenblatt unveiled the Perceptron, a device that could learn to classify images by adjusting weights on its inputs. The New York Times reported the Navy expected it to walk, talk, see, write, reproduce itself, and be conscious of its existence.

The actual hardware, the Mark I Perceptron, was more modest. It was a room-sized analog machine that could learn simple visual distinctions, like telling triangles from squares. Still, it worked, and the idea of a self-modifying machine captured imaginations.

The Minsky and Papert critique

In 1969, Marvin Minsky and Seymour Papert published Perceptrons, a rigorous book that proved single-layer perceptrons could not learn the XOR function or any non-linearly-separable pattern. The math was correct. The message was often read as: neural networks are a dead end.

• Single-layer perceptrons are limited to linear decision boundaries
• Multi-layer versions would lift the limit, but no one knew how to train them
• Funding for neural approaches largely dried up for over a decade
• Rosenblatt died in 1971; some say the field turned its back on him

“Perceptrons have been widely publicized as pattern-recognition or learning machines. Most of this writing is without scientific value.”

The big idea: a powerful critique at the wrong moment can freeze a field for years. The perceptron was not wrong; it was just incomplete, and nobody had the tools to finish it yet.