Floor 6 — The Perceptron and the Winter That Didn't Have To Happen

The floor

In 1958, Frank Rosenblatt — a psychologist at the Cornell Aeronautical Laboratory — unveiled the Mark I Perceptron. A physical machine, four hundred photocells connected to motor-driven potentiometers, wired to a single output neuron. Show it images. Tell it when it was right and when it was wrong. The potentiometers turned. After enough examples, it recognised letters.

It learned. Not metaphorically. Physically, with servomotors and copper.

The New York Times ran a piece quoting the Navy: the perceptron was "the embryo of an electronic computer that... will be able to walk, talk, see, write, reproduce itself and be conscious of its existence." The hype was absurd. The device was real.

Eleven years later, Marvin Minsky and Seymour Papert published a book called Perceptrons. It proved, rigorously, that a single-layer perceptron could not compute the exclusive-or function.

The field collapsed.

What was picked

The narrow reading. Funders read "perceptrons cannot compute XOR" and heard "neural networks don't work." Money moved to symbolic AI and expert systems. Graduate students changed topics. The connectionist research program went into hibernation for roughly fifteen years. This is what the history books call the first AI Winter.

The narrow reading was not wrong, exactly. Single-layer perceptrons really cannot do XOR. Minsky and Papert's proofs were correct.

The narrow reading was only half the book, though.

What could have been picked

Rosenblatt had already been working on multi-layer perceptrons. The math for training them was not available in closed form, but he, and others, understood that adding a hidden layer fixed the problem that Minsky and Papert identified. Backpropagation — the algorithm that makes multi-layer training work — had already been published in the context of control theory and statistics. Seppo Linnainmaa wrote it down in his master's thesis in 1970. Paul Werbos re-derived it in 1974. It would be re-discovered again in 1986.

The mathematics of the alternate timeline was on bookshelves the whole time the winter lasted.

The fork was not taken because the proof was persuasive. The fork was taken because the proof gave the symbolic AI community the opening they needed to argue their approach was the serious one. Minsky was part of that community. The book's final chapters, which gestured at how multi-layer networks might be harder but not impossible, were read much less than the middle chapters where the restriction bit.

A book can close a door that the mathematics never shut.

What we missed

A decade. Maybe more.

Frank Rosenblatt died in 1971, age 43, in a sailing accident on Chesapeake Bay, on his birthday. We do not know what he would have done with the next forty years of his career. We know the field went into a dark room and did not come back out until the mid-80s. We know the students who would have worked with him went elsewhere.

This floor has a kind of grief that the earlier floors do not. The others are grief for a road not taken. This one is grief for a road actively blocked.

The alternate timeline is not a technical one. It is a sociological one. A field where Minsky and Papert published the book as a challenge rather than a verdict. A field where Rosenblatt's work was extended rather than buried. A field where the word "neural" was not embarrassing to use at a conference from 1972 to 1985.

What the next floor will ask

If the connectionist door was closed, what walked through the one the field opened instead?

That's Floor 7. It involves a lot of rules.