1. Feasibility
In January 1957, a company in Buffalo, New York, called the Cornell Aeronautical Laboratory printed a numbered report.
The title on the cover was “The Perceptron: A Perceiving and Recognizing Automaton.” There was one name on the cover, Frank Rosenblatt, and a job title, project engineer. Two more names were on the title page, the people who had approved the work. The whole thing ran to fifty-eight pages. A cover, a number, a job title, two approvals. This is what something looks like when it is meant to go on file: people will turn its pages, check what is there, and use it to decide where the next money goes.
There are two verbs in the title: perceive, recognize. Put something in front of it; first see it, then say what it is. That is the problem the document sets itself. No thinking in the title, no understanding, no intelligence. Both verbs are about what to do once a picture comes in.
Not far into the text, a sentence gives the aim of the present stage: establish the technical and economic feasibility. Feasibility was not a scientific term that year. It was something people in engineering and business said every day. Can it be done, and would it be worth doing? Put a small sum into finding out, then write up what you found. A document like that does not announce that the thing has been done. It says it is worth looking into further.
There are two kinds of feasibility, not one. Technical feasibility asks whether it can be made. Economic feasibility asks what it will cost and whether it will be worth the money. People rarely mention the second half afterward. This sort of document has a definite reader: somebody deciding whether to pay for the next step.
The qualification comes from the man working on the project. Everything that happened later went a little farther from that sentence. The sentence stayed where it was, at the start of the fifty-eight pages. Nobody moved it.
The later story flattened it: a university, a professor, a machine that could learn. On the paper at the time there was a company, a project engineer, a numbered report, and two signatures approving it. The flat version is easier to tell because it has one person in it. The other version has a process, and a process is harder to tell.
The January report says the laboratory's internal research program paid for it. Not until July 1957 did the Office of Naval Research begin funding the project. Frank Rosenblatt was still at the company in Buffalo. He would move to the university in Ithaca a few years later.
2. Washington, July 7
Washington, July 7, 1958. That day the Navy showed the press something. On page twenty-five the following day, the New York Times printed a wire-service report datelined Washington, July 7. The headline was “New Navy Device Learns by Doing.”
The report identifies the machine put in front of the reporters. It was a large computer borrowed from the Weather Bureau, made by a commercial machine company, about five tons in weight, and big enough to occupy a room. At this point the perceptron was a method being simulated on a general-purpose computer, not a special-purpose machine that had been built.
An institution shows reporters its new device. What is on display is another department's general-purpose computer. The report that day says so plainly. Later the detail disappears, because there is nowhere for it to go in the story.
The report says exactly what the machine was asked to do: tell whether a mark on a card lay on the left or on the right. After fifty tries it could tell. That was it. No letters, no faces, no cats or dogs. Left or right, fifty tries.
It is a small thing to say: tell left from right. What makes it less small is that in those fifty tries nobody changed the program. What changed was a group of numbers in the machine itself, in response to its own wrong answers earlier on. That was what the roomful of reporters saw.
There is a sentence in the report about the future. It would be quoted for more than sixty years. In the edition that came out the next day, the subject of the sentence was not Rosenblatt. The Navy was the one expecting something. The Navy expected the machine to walk, talk, see, write, reproduce itself, and be conscious of its existence. The expectation is attached to the Navy. The list is not in quotation marks or put in the mouth of an individual.
In Chinese, a familiar later version goes like this. Rosenblatt said his perceptron would soon walk, talk, see, write, reproduce itself, and be self-aware. The list is the same. The subject is different. The news report separately gives a conjecture of Rosenblatt's own: he thought there might in principle be a route toward self-reproduction and consciousness. That is not the same statement as the whole list of expectations in the Navy's name.
The Navy expects: an institution paying for a project speaks about the future. Expects leaves a little room. In this report no individual claims the whole list in a quotation. A person says is different. Then it is a person's word for something in their own hands. The statement can be traced to that person, and the person has left other writing to compare it with.
Change the first sentence into the second and nothing is added to the list. The subject changes. Over the next sixty years, when somebody wanted an example of a scientist making grand claims, this was often the sentence that came out.
The same thing in the same period went by three names, spoken by three different voices. The newspaper called it a new Navy device. Rosenblatt's report from 1957 called it a perceiving and recognizing automaton. His paper in Psychological Review in November 1958 called it a probabilistic model for information storage and organization in the brain.
Three names, three distances. The first describes a piece of equipment. The second describes a machine that does two jobs. The third describes a model of the brain. What survived of the three was a variant of the first.
A Boston paper ran the story the same day under a headline saying the Navy was going to make a thinking robot. The same day, the same wire report, two headlines. One says learning by doing, the other a thinking robot. Put the two next to each other and you can already see which way the news is going.
Five months after July, on December 6, 1958, the New Yorker published a short piece called “The Rival.” In that account the physical perceptron had still not been built. What had been done was a computer simulation. It had no ready practical use. Five months apart, the later account makes a narrower claim. Both accounts were in print at the time.
A roomful of people, a five-ton computer, a pile of cards. The whole thing that happened in the room was that a machine learned to tell left from right within fifty tries. The news from the room was that a machine would one day know it existed. The two sentences are about the same afternoon.
3. Four Hundred Points of Light
On February 15, 1960, the same company issued the Mark I Perceptron Operators' Manual. By then they had built the machine.
On the front is a twenty-by-twenty light-sensitive grid. Four hundred cells, each a point that senses light. Put a card in front of the grid, and the black and white on the card become light and dark at those four hundred points.
The manual divides the parts into three kinds. First are the four hundred light-sensitive points. Second are a group of intermediate units, five hundred and twelve of them. Third are eight output units.
Light strikes a point and it sends twenty-four volts onward. An intermediate unit receives signals like that along several lines and adds them up. Reach a given number and the unit activates; fall short and it does not. The manual calls that number the threshold. It can be adjusted.
There are eight outputs, not one. The machine can give more than a yes or no. It can put the card in front of it into one of eight categories.
The four hundred light-sensitive points are connected to the five hundred and twelve intermediate units. These connections are fixed; once installed, they do not move. An intermediate unit can receive up to forty points as inputs. It sees only some of the points on the grid.
This layer stays put. Nothing the machine learns later takes place in it. The machine cannot change where it looks. It can only change whom it listens to. The book that comes later will reckon with this layer.
What can be changed is in the layer after it. The amount that an intermediate unit sends to the output can be adjusted. The manual calls the adjustable quantity a weight. On this machine a weight is a potentiometer.
A potentiometer is like a knob on a radio. Turn it a little and more or less current gets through. The knobs on the Mark I are not turned by hand. Little motors turn them. During training, voltage from the output side feeds back to drive the motors. A motor turns the knob through an angle.
There is a loop. The answer comes out at the output. Whether it was right or wrong comes back from that end as a voltage, then as a movement of the motor, then as a new position of the knob. When the next card comes in, the machine is not quite the one that was there a moment before. For this machine to learn once is for a group of knobs each to move a little.
Here the word learning has a shape you can see. Motors move, shafts turn, knobs take up new positions. This is no metaphor. What a machine had learned in 1960 was a set of physical positions. You could measure them with a ruler or feel them with a hand. Turn off the power and there they stayed.
Put the two layers together and this is how the machine does its work. The intermediate units each look at some of the points and give a result. The knobs decide how much each result counts on the way to the output. The output units add what they receive and compare the sum with a threshold.
Training changes how loud the voices are. After a wrong answer, the training signal can turn up some of the contributions that go to the right output or turn down some that go to the wrong one. The 1969 book asks a question that grows out of this arrangement. When a group of units each see some of the points and their results are added together, what can they decide?
There are somewhere between a few hundred and a few thousand things you can turn in the whole machine. A model in a phone today can have more than a million times as many things to adjust.
The machine survives. The National Museum of American History keeps two cabinets, each taller than a person. There is a photograph in the manual. Everything in it is still: cabinets, panels, wires.
Later the Mark I learned to tell a square from a circle, a square from a diamond, and some groups of letters apart. It could learn to do those things.
At this point one can say what the machine looks like. Four hundred points of light, a layer of fixed wiring, a set of intermediate units that each see only part of the picture, a layer of motor-driven knobs, eight outputs. This shape has a boundary. Nine years later somebody took its measure.
4. What the Book Proved
In 1969 MIT Press published a book, Perceptrons: An Introduction to Computational Geometry. The text ran to 258 pages. It had two authors, Marvin Minsky and Seymour Papert.
Minsky's name has been in the earlier essays. It is on the paper asking for money for the summer of 1956. He wrote the book on computation in 1967. His was also the name printed over the direct quotation about three to eight years in 1970. In 1969 he was a co-director of the artificial intelligence laboratory at MIT. Papert was already a professor of applied mathematics there and later served as a co-director of the lab too.
The title is plural: Perceptrons. That plural has been singled out since. The mathematical object is not one machine but a class of machines. The book defines its shape. At the front are units that each see only some of the points in an image and give an answer of yes or no. After them the answers are added using their respective coefficients and put through a threshold to get an overall answer. The shape corresponds to the Mark I: four hundred points of light in front, units that each look at some points in the middle, adjustable coefficients after them, output at the end.
The first thing a mathematics book has to do before proving anything is pin down the object of the proof. Once it is fixed, there is something to calculate. In the moment the object is fixed, the scope of the book is fixed too. Much of the later dispute turns on this. Is the class pinned down on the page the same as the broad territory the readers have in mind?
The question is what a machine of this shape can and cannot decide. To ask it, the book assigns a number to each kind of decision. At the very least, how many points must the unit that sees the most be allowed to inspect at once? If seeing one is enough, the number is one. If at least three must be seen together, it is three. The book calls this the order.
What is useful about order is that you can follow it as the image grows. Go from ten points to a hundred or a thousand and see whether the order required to decide the question stays put or goes up. If it goes on rising, you cannot finish the job with a fixed upper limit for the number of input points each unit can see. Order in the book is about the number of points. It does not require them to be next to one another in the image.
The person who built the machine knew his could tell a square from a circle. Whether the shape of it was sufficient in itself remained open. Two people took several years to work that out. They came up with a set of results.
Putting a limit on the number of points a unit can see has a counterpart in the actual device. The wiring specified in the manual allows one such unit to be connected to at most forty light-sensitive points. In 1960, seeing all four hundred at once would have meant bringing four hundred wires to the same place. The machine was designed precisely so that this would not be necessary.
The class defined in the book and the machine made in Buffalo can be compared on that point. Whether the class is the same as the whole line of research is a different question. Someone posed it the following year.
One chapter is about parity. Is the number of points alight in a picture odd or even? The order grows as the image grows, with no upper bound. To decide odd or even, some unit must see the whole image at once. What is being asked here is what happens when the number of points any one unit can see is capped, but the image can keep getting bigger.
Two points can light up in four ways. It is odd if exactly one is lit. This is exclusive OR. If you try to decide it linearly from those two inputs directly, you can draw the four cases as four points, but no straight line separates the odd from the even. Let one unit look at both inputs first and it is no longer a question of drawing a line through those four points. The book's argument is more general than the little example. It asks how many points need to be seen at the same time.
Another chapter is about connectedness. Is a shape joined in a single piece or broken in two? The book calls it a geometric property whose order has no upper bound. Look at two simple pictures and a person can often tell at a glance.
Think about it in terms of the earlier machine. The five hundred and twelve units crouch behind the grid, each looking at the few points wired to it. Put a figure in front of them and the break could be anywhere. A unit that happens to see the points near the break might be able to call it out. To other units the picture with a break and the picture without one look exactly alike.
Guessing which part a unit might miss in one picture is not a proof. What the book says is that as the figures grow, there is no fixed limit on the number of points per unit that will let this method of adding local judgments always get connectedness right.
As long as the number a unit can see has a fixed upper limit and the image can be made arbitrarily large, there is no way to decide it, whatever coefficients the parts are given. Something a person can often see immediately in simple figures is something a class of machines cannot decide on account of its shape. That is the boundary they measured.
Other chapters deal with diameter-limited perceptrons, the size of coefficients, learning, and linear separation and learning. Two headings include learning. So the book does not pass over learning either. Turn 258 pages into a proof about four points and most of the text drops out.
5. Or Refute It
Chapter 13 is the last. It asks what would happen if the earlier structure were extended to several layers.
There is a sentence on page 231 that is quoted often, usually only its latter part. The whole sentence is a request. The authors ask later researchers to elucidate (or reject) “our intuitive judgment that the extension is sterile.” The words or reject are an aside in parentheses in the original.
Now shave the sentence down one layer at a time.
Take out the parenthesis: “Our intuitive judgment” is that the extension is sterile.
Take out our: the intuitive judgment is that the extension is sterile.
Take out intuitive judgment: the extension is sterile.
Take out the particular extension: multilayer approaches are sterile.
Go a step farther and put it another way: this book proved that multilayer approaches do not work.
The first sentence is a request for research. The last sounds like a theorem. The cuts above are a way to set the original and the usual retelling alongside one another. They are not a sequence of transmission that has been recovered. What stands out between the two ends is how many qualifications are gone.
With each layer gone, the sentence becomes a little more useful. A request is hard to wield as a quotation. A theorem will do. An intuitive judgment cannot be used to close a case. A theorem can.
Strip it down all the way and you no longer need the book's title. “It has been proved not to work” is enough. Any book can hang from the end of it. When a sentence is shaved down that far, it no longer belongs to the person who wrote it. It belongs to anybody who needs it.
The proofs in the book have a recognizable shape. First comes a statement, with a number. Beneath it comes the proof, and there is an ending when the proof is complete. There are dozens of these in 258 pages.
The sentence on 231 is not numbered. It has no proof under it. It is part of a passage that asks others to do work. Open the 1969 book and the layout tells the two apart at a glance. A retelling does not carry the layout. When the words leave the paper, the number, the proof, and its ending stay behind. All that travels is what the sentence says.
In the same passage, the authors leave two more things to the future: a convergence theorem that might exist, and deeper reasons why learning might fail. The grammatical shape of the passage is that of a research question, not an impossibility already proved.
Two signed reviews appeared after the book came out. Allen Newell's appeared in Science on August 22, 1969. His name is in the preceding essay. He and Simon were jointly responsible for the four predictions read out at the dinner in 1957. Newell's review rates the book very highly. A general science weekly printed a very favorable review in the summer of the year the book appeared. Not a judgment in court, not a resolution or an announcement. A review.
The other review was by Block, in Information and Control in December 1970. He thought the class of machines analyzed was too restricted, that it could not simply be equated with Rosenblatt's research program as a whole, and that the title was misleading on that account.
A book called Perceptrons had done the mathematics for a class it defined itself. Someone said so openly in December 1970. He was someone in the field, and he said it in a journal for his colleagues, over his own name.
The two responses coexist, sixteen months apart: very high praise and an objection about scope. One book was published and two reviews followed. A book crushing a field is the shape the story took later. In the following sixty years Block's objection would be made again many times. It was already on paper in December 1970.
6. The Epilogue Twenty Years Later
In 1988 the same publisher brought out an expanded edition. An epilogue titled “The New Connectionism” ran to thirty-four pages after a book from nineteen years earlier. When a book needs an epilogue to explain what it said at the time, it is no longer wholly under its authors' control.
They wrote the epilogue because they had something to answer. In 1986 David Rumelhart and James McClelland edited two volumes called Parallel Distributed Processing. The books marked the way this line of work was coming together again.
A passage in the first volume said Minsky and Papert's appraisal of perceptrons had wrongly implicated other networks that were more interesting and more powerful. Implicated does not say the book got the mathematics wrong. It says the reputation of the book fell on things it had not analyzed.
The 1988 epilogue reproduces that passage in full, then says that the authors can hardly recognize themselves in the description.
They then draw a boundary of their own. They say most of the original theorems were explicitly about machines with just one layer of adjustable connection weights. In 1988 this is how the authors qualify their 1969 book. It points in the direction of Block's 1970 review. After eighteen years, the point moves from outside the book to inside it, from the reviewer to the authors themselves.
They immediately add that this does not mean the phenomena are automatically irrelevant to multilayer cases. You have to look at each case. The epilogue splits the problem of the time in two. A machine with a single layer of adjustable weights has a learning procedure that converges, but its powers of representation are limited. A multilayer net can represent much more, but had long lacked an equally reliable method for learning. The authors go on to examine the rule for adjusting the weights that had been proposed by then.
Over the following pages they reject presenting gradient and hill-climbing methods as universally effective. They raise issues of scale, local minima, the kinds of problems, and structure. At the same time they explicitly favor using the methods empirically in new fields, and expect such work to yield important progress. The epilogue does not stop at a no.
The epilogue says what boundary the authors accepted twenty years later. What was written in 1969 can still be found on the paper of 1969. The parentheses on page 231 were there then. The two sides of the exchange twenty years later are not talking about the same thing. One asks what the book was used for afterward. The other asks what it said at the time.
7. The Papers of Those Ten Years
On July 11, 1971, Rosenblatt died in a boating accident on Chesapeake Bay. It was his forty-third birthday.
The usual story about the decade after the book goes like this. This line of work died. No money, no people, no papers. In the case of the papers, the printed record is there.
In 1972 Shun-ichi Amari's “Learning Patterns and Pattern Sequences by Self-Organizing Nets of Threshold Elements” appeared in a computing transactions journal. The same year, Teuvo Kohonen's “Correlation Matrix Memories” ran in the same journal, and James Anderson's “A Simple Neural Network Generating an Interactive Memory” appeared in a journal of mathematical biosciences. In 1973 Christoph von der Malsburg published “Self-Organization of Orientation Sensitive Cells in the Striate Cortex” in a journal of cybernetics. In 1975 Kunihiko Fukushima published “Cognitron: A Self-Organizing Multilayered Neural Network” in a journal of biological cybernetics. In 1976 Stephen Grossberg published “Adaptive Pattern Classification and Universal Recoding.” In 1978 Geoffrey Hinton submitted his Edinburgh dissertation, “Relaxation and Its Role in Vision.”
These are the words of those years: self-organization, threshold elements, correlation matrices, recoding, relaxation. The researchers were doing their own work in their own disciplines. Some were in physiology, some in vision, some in memory. They carried on with their subjects, which in the first place did not belong to AI. For a line of work to be crowded out of one field does not mean it has been crowded out of them all.
The 1975 paper already has multilayered neural network in its title. Six years after the book that supposedly proved layers could not work, a paper with multilayered in the title appeared. Both things belong to the record. Against the claim that there was not a paper for ten years after 1969, there is the list above.
In 1996 Mikel Olazaran published a study of the dispute in a sociology of science journal. In his account, as early as the late 1950s and early 1960s, the competition between the perceptron approach and its rival was entangled with professionalization, structures of authority, and allocation of resources. After the book appeared, work in this direction did decline markedly and was pushed into disciplines outside artificial intelligence. A few people still did it. His claim is not that nothing happened. The decline and the move happened together.
He did something else too. He made the version in which the field died when the book came out into a subject for study, calling it an official history.
He was the one who went looking in different places for some of the names above and collected them in one place: a group in Europe, one in America, one in Japan. The papers had appeared in journals at the time, where anyone could look them up. Putting them together as counterevidence had to wait twenty years, until somebody asked how this bit of history came to be told as it is now. A thing not remembered is not the same as a thing that did not happen. The distinction came up in the previous essay.
The two volumes of 1986 spoke of networks being implicated. In a retrospective in 2019 the university itself used stronger language, saying the book sealed the fate of Rosenblatt's work. From implicated to sealed, thirty-three years. The two passages cannot give us an account of what funding fell, when, or by how much. Their wording can be compared. The latter is heavier.
8. Two Changes in the Telling
There are two retellings in this essay. The first starts in 1958. A statement about the future ran in the newspaper under the Navy's name. The Navy expected it would one day do the things in the list. In the version that came down, the list was in one person's mouth: he said the machine he had made would do those things before long. Not one item in the list changed. What changed was the first half of the sentence.
An institution's expectation and a person's promise carry different responsibilities. When a wire report records an expectation on behalf of an institution, the question is what the institution said. When someone promises something about a device in his hands, the first question is whether those were his words. What has been cut away is the layer of attribution. Once it goes, the sentence is in one person's name.
The second starts after 1969. The sentence was originally a request. The writers offer an intuitive judgment and ask people who come after them to make it clear (or reject it). The judgment is that the extension will be sterile.
The later sentence sounds like a theorem: the book proved multilayer approaches could not work. Three things have been dropped: or reject, our, intuitive judgment. One is a question and the other is a conclusion. Between them lies the work the authors asked someone else to do.
The changes go in opposite directions. The first makes the statement bigger. The second makes it final. They are the same move: take away the qualifications.
Qualifications are how speakers leave themselves room. They take up words in the original and are the first things squeezed out in the retelling, because they do not carry the story. A story wants a subject, an action, a result. The Navy expected does not give it that shape. A man bragged does. Please go ahead and overturn our intuitive judgment does not. The book proved it could not work does.
Put the two results together and you get the story now told. A man made a grand claim. A book proved him wrong. The line of work died for ten years as a result. Each sentence is what remains when a statement with qualifications loses them.
The first put the expectation in the Navy's name, not his. The second left a judgment for later people to overturn, not a result already proved. The word behind the third is implicated, not frozen. And there are papers from those years, one of them with multilayered in the title.
There is another thing here that differs from the preceding essays and deserves to be set out on its own. The 1969 book measures a boundary. Most of the boundaries earlier were drawn in sentences: a machine can only do what it is told; a machine cannot originate. Later people can repeat or dispute them. They have no dimensions.
This time, in 1969, the book sets the machine's shape down in a definition: units that can see at most so many points each, a layer of adjustable coefficients, a threshold. It then asks what this shape cannot decide. The answer comes with a number that can be worked out. How many points must some unit be able to see at once for this thing to be decided?
For parity that has to be the whole image. For connectedness there is no upper bound either. This time the part left over has been measured.
A measured boundary can be checked. Copy out the definition and do the calculation. To go farther, let a unit see more points at once, or stack layers to combine the judgments again. Once the shape changes, you have to ask anew what it can do.
A thing that has been built has a shape, and there are always things left outside the shape. In the earlier essays the part left over is called the remainder, and usually one can only point to it and say there is more. In 1969 the remainder was written in dimensions: how many points, whether the number grows, whether there is a limit.
What travels afterward is a verdict. The dimensions get reduced in the retelling to “It cannot be done.” It goes that smoothly because dimensions come with qualifications and a verdict travels without them. These machines, this shape, this kind of locality, this threshold, the way the number grows with the image: the measurement carries them all. The verdict carries none.
The more precisely you measure a boundary, the more qualifications it has. The more qualifications it has, the less readily it travels. There is a difficulty here. You measure the boundary clearly so someone can go on calculating. The more clearly you measure it, the more it loses as it travels. A book that gives no dimensions goes out into the world as a sentence. A book that gives dimensions goes out as a sentence too, and the sentence bears the title of the book.
You can see it here because the documents at both ends are still there. The aside is still on the page from 1969. The subject of the sentence is still on page twenty-five for July 8, 1958.
There are three gaps in this essay, none at the edges. We have not found when the sentence moved from the Navy's name to the name of a person. What can be established are the two ends: the newspaper puts the six expectations in the Navy's name, while some accounts now put them in a person's. We have not found when a book went from a request to a verdict. Page 231 in 1969 leaves a request. Today what travels is a theorem. Nor has a full account been found here of the money this line of work received in America over those years, or of who first wrote that one book froze funding for a decade.
Three gaps and three things people usually say. Each has only one end left. The end that has survived is the one that is easy to tell.
Back, finally, to the numbered report of January 1957. Its sentence about establishing feasibility has two qualifications, technical and economic, and it is written by the person doing the work.
A year and a half later, on the newspaper page the same thing is going to walk, talk, see, write, reproduce itself, and know that it exists. Eleven years on, a book measures what the same thing cannot decide. On page 231 the book leaves a parenthesis. Farther on the parenthesis is gone and feasibility is gone too. Two lines are left: too much was claimed for it, and later it was proved not to work.
Neither is what was written. Both are short, tidy, and without qualifications. Both are useful. The originals have stayed on the page for more than sixty years without moving.
The machines are still being built. Each generation says again where the boundary lies.