Perceptrons (Minsky and Papert): What one layer cannot compute

The event and the narrow claim

In 1969 the MIT Press published Perceptrons: An Introduction to Computational Geometry, by Marvin Minsky and Seymour Papert of MIT's Artificial Intelligence Laboratory. Allen Newell of Carnegie-Mellon University reviewed it in Science on 22 August 1969 under the title "A step toward the understanding of information processes", and the review's header records a slim volume of 258 pages, priced at $12 in cloth and $4.95 in paper ( Newell's review in Science). The book is a sequence of theorems about what one class of learning machines, the perceptrons that Frank Rosenblatt had introduced in 1958, can and cannot compute when each of their feature detectors is limited in how much of the input it can see.

Two claims about the book need to be kept apart. The narrow one is mathematical and not in dispute: a single layer of adjustable weights sitting on top of local, fixed feature detectors cannot recognize certain global properties of an image, such as whether the number of lit points is odd or even (parity) or whether a figure is all in one piece (connectedness), unless some detector looks at nearly everything. The broad one is historical: that the book persuaded the field and its funders to abandon neural networks for more than a decade. That second claim is widely repeated, including in Cornell's 2019 account of Rosenblatt's career, which says the book essentially sealed the perceptron's fate. Historians and participants have contested it for more than thirty years. This report sets out what the book proved, what its authors guessed, and why the causal story is less settled than its fame suggests.

A rivalry older than the proofs

The book was the last move in an argument that had begun a decade earlier. Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, described the perceptron as a probabilistic model of information storage in the brain in 1958, and the Office of Naval Research funded both the theory and a hardware machine (see The perceptron). By March 1961 his long report Principles of Neurodynamics had already named his main critic: a footnote recommends Minsky's essay "Steps Toward Artificial Intelligence" as an entertaining statement of the views of "the loyal opposition". The same report conceded that two problems, separating a figure from its background and recognizing topological relations, were new territory where few inroads had been made.

Seymour Papert, a mathematician who had worked with Jean Piaget in Geneva, joined MIT as a research associate in 1963 and later became co-director of the AI Laboratory under Minsky, according to MIT's obituary of Papert. The sociologist Mikel Olazaran, who interviewed many of the participants for his 1991 Edinburgh doctoral thesis, places the decision to pin down the perceptron's limits at about this time. Minsky told him that by 1965 people were worried, because perceptron groups wanted money for bigger machines that did not seem to be going anywhere. Papert later acknowledged, in a 1988 essay the thesis quotes, "some hostility" behind the work and a wish to stop funding being spent on what he considered misleading applications. The book's own introduction, also reproduced there, speaks of its authors' "therapeutic compulsion" to dispel a holistic misconception about loosely wired, brain-like machines, and estimates that perhaps a hundred groups had experimented with Rosenblatt's model.

None of this makes the mathematics wrong. It does mean the book was written as an intervention in a contest for money and legitimacy, and that is the frame its later readers inherited.

What the theorems actually establish

The core results were in print two years before the book. A 1967 paper in the American Mathematical Society's Proceedings of Symposia in Applied Mathematics, circulated as MIT AI Memo 140, "Linearly Unrecognizable Patterns", lays out the framework. A pattern is a set of lit points on a retina. A perceptron decides a property by adding up weighted votes from simpler predicates and comparing the total with a threshold. The key quantity is order: the smallest number k such that the property can be decided this way when each simpler predicate depends on at most k points. The memo's research was supported through Project MAC by ARPA under Office of Naval Research contracts, the same office that funded Rosenblatt. Its examples are concrete. Convexity has order at most 3, because a figure is convex exactly when no pair of its points has a midpoint outside it, and each such test involves three points. The property "the figure contains exactly M points" has order 2.

The memo also shows the proof technique that made the book famous. For the smallest case of parity, two inputs, which is exclusive-or or its complement, it observes that the property does not change if the two inputs are swapped, so the two weights can be averaged and treated as equal. The weighted sum then depends only on how many points are lit, and a linear function of that count cannot be true at zero, false at one and true again at two, because it would have to change direction twice. The general version, the group invariance theorem, lets the authors replace any perceptron for a symmetric property with one whose weights depend only on classes of equivalent detectors. In the book, this line of reasoning shows that the order of parity is the whole retina: at least one detector must look at every point, as Olazaran's thesis summarizes the parity chapter.

Connectedness is the showpiece. The 1967 memo states that there is no k for which connectedness is of order k, and both memo and book treat a diameter-limited variant, in which each detector sees only a small patch, that also fails. The book's own illustration, quoted in the thesis, is that a perceptron for connectedness on a 100 by 100 toroidal retina needs partial functions that each examine many hundreds of points, so the supposedly local parts are themselves global. The chapter list of the MIT Press edition shows the architecture of the argument: linear Boolean inequalities, group invariance, parity, connectedness as a geometric property with unbounded order, then chapters on the size of coefficients, on learning, and on pattern recognition.

Two features are easy to miss. First, the limitation concerns locality, not computability. The memo's positive normal form theorem shows that every property can be decided by a threshold unit if detectors may examine any subset of points; the cost is that the order then grows with the retina. Second, the authors contrasted parallel and serial machines and noted, in a passage Olazaran cites, that every pattern they discussed is easy for a simple program on an ordinary computer.

Why XOR became the shorthand

The popular version reduces all of this to one fact: a perceptron cannot learn exclusive-or. That is true of a single threshold unit reading the two raw inputs, it was understood before 1969 because exclusive-or is not linearly separable, and it is the smallest instance of parity, which is why it became the teaching example. But it misstates the book's subject. The book asked how the required order grows as the retina grows, and its targets were global geometric properties rather than a two-bit truth table.

The shorthand also hides the book's positive results. Papert claimed in 1988 that more than half of it was devoted to "properceptron" findings about surprising things perceptrons can do, a claim Olazaran records alongside his observation that, by then, the authors wanted to loosen the link between their names and the eclipse of neural networks. Léon Bottou, rereading the book for a 2015 AAAI symposium talk, lists among its results that the Euler number, a topological count of components minus holes, has low order, while parity and connectedness do not. Bottou called the order analysis brilliant and pointed out that it also describes what a convolutional network with a final pooling layer can compute, a reminder that the framework is not a museum piece.

Nor were the limits news to the perceptron camp. Rosenblatt's preface to Principles of Neurodynamics complained that the word perceptron had been misread as one piece of hardware when he meant a family of theoretical nerve nets, and the same volume contains chapters on three-layer and four-layer systems and a section on what he called a back-propagating error correction procedure. That procedure assigned blame to hidden units probabilistically. It was not the gradient method later known as backpropagation, but it shows that training more than one layer was an acknowledged open problem inside the program the book criticized.

The guess about many layers

The book's most consequential sentence was not a theorem. In its closing chapter on perceptrons and pattern recognition, the authors argued that none of the single-layer perceptron's virtues (its linearity, its learning theorem, its simplicity) need carry over to the many-layered version, and called it an important problem to elucidate or reject "our intuitive judgment that the extension is sterile". Quote Investigator's check against scans of the 1969 text confirms the wording and its place on pages 231 to 232. The judgment was explicitly labeled intuitive, and the passage invited someone to discover a convergence theorem for multilayer machines.

Contemporaries on the other side read it as a misframing. H. D. Block, Rosenblatt's Cornell colleague, argued in a long 1970 review in Information and Control that the simple perceptron was what enthusiasts studied first, not what they considered typical, and that Minsky and Papert had treated a temporary handhold as the final structure. He added that people cannot see the parity of a large set of dots either, or tell at a glance which of the two figures on the book's cover is connected, so failing at those tasks hardly disqualified a model of human perception. Bernard Widrow, whose Stanford group built the adaline, told Olazaran that the authors had "defined the idea of a perceptron sufficiently narrowly" to prove it could do nothing, while calling the book good mathematics. Both responses are documented in Olazaran's thesis.

The guess also met an early test. Paul Werbos told Olazaran that in 1970 or 1971 he took Minsky a model of intelligence that included a gradient method for training hidden units, and received a discouraging reply. Minsky, interviewed separately, remembered the idea as workable in principle but impractical: slow to converge, prone to getting stuck in local minima, and costly on the computers of the time. Werbos later wrote that, because neural network applications were not considered acceptable, his 1974 thesis framed the method around conventional forecasting.

Did the book kill neural networks?

The strong claim has an authoritative-sounding witness in Minsky himself. Speaking to the journalist Jeremy Bernstein for a 1981 New Yorker profile, quoted in Olazaran's thesis, he said there had been several thousand papers on perceptrons up to 1969, "but our book put a stop to those", and joked that the book had been too good, leaving students nothing to do. Stephen Grossberg, a professor at MIT from 1967 to 1975, later described the dampening effect of what he saw as the authors' hostility to neural network research. In 1988 Minsky was reported to have regretted the book's chilling effect.

The evidence for a different account is at least as strong. Olazaran's 1996 article in Social Studies of Science argues that the official reading of the debate, that Minsky and Papert had shown neural networks to be a dead end, was a product of the emergence, institutionalization and legitimation of symbolic AI, with its own resource allocation and authority structure, and that at the level of working researchers there was considerable interpretative flexibility. In the thesis he argues that problems such as parity became decisive anomalies as a result of the controversy's closure, not as its cause. His interviews show the major groups changing course before the book appeared. Nils Nilsson recalled that about 1965 or 1966 the neural network group at the Stanford Research Institute decided it had gone about as far as it could; the group turned to symbolic vision and planning and in 1966 began the Shakey robot. Widrow's group moved to adaptive antennas and filters, where its least-mean-squares rule found lasting use in modems and echo cancellers.

Money mattered as much as mathematics. Nilsson, in The Quest for Artificial Intelligence, doubts that the book caused the fading interest, noting that Rosenblatt himself had begun concentrating on other topics well before 1969 and that the success of heuristic programming drew researchers, Nilsson included, away from networks in the mid-1960s. He also records that on 19 November 1969 Congress passed the Mansfield Amendment, which confined Defense Department support for basic research to projects with a direct and apparent military function. At a July 1971 tribute quoted by Olazaran, the Cornell biologist Richard O'Brien called Rosenblatt a victim of that amendment whose grants had "melted like summer snow". Olazaran's sources add that Rosenblatt never stopped building perceptrons, so Nilsson's point is about divided attention, not abandonment. Minsky, for his part, told Olazaran that the story of DARPA refusing neural network funding was a myth, because there were no good ideas to fund. Every one of these witnesses had a stake in the outcome.

What can be said with confidence is narrower than either camp's version. The central arguments were in print by 1967, two of the leading groups had moved on by 1966, defense funding was tightening, and Rosenblatt, the program's most energetic advocate, drowned in a sailing accident in 1971, as Cornell's account notes. The book was the most authoritative statement of a judgment the field was already reaching, and it gave that judgment a rigorous form that outsiders could cite. It belongs to the same climate as Britain's Lighthill report, but it was neither the sole nor clearly the main cause of the decline.

The 1988 rejoinder

When neural networks returned, they returned through the book's open question. In 1986 Rumelhart, Hinton and Williams showed in Nature that back-propagating errors through layers of hidden units lets a network learn its own internal features, and their abstract contrasts this with the perceptron-convergence procedure (see Learning by backpropagation). In their longer chapter in the Parallel Distributed Processing volumes they reprinted the sterile passage and wrote that they had answered Minsky and Papert's challenge, as both Quote Investigator and Olazaran document.

Minsky and Papert replied with an expanded edition in 1988. Its new prologue says they found little of significance had changed since 1969, so they kept the original text with its 1972 corrections and added an epilogue, and it offers their version of the history: progress "had already come to a virtual halt because of the lack of adequate basic theories", as Quote Investigator quotes it. The epilogue, "The New Connectionism", runs to more than thirty pages in the MIT Press edition. It calls backpropagation merely a way of computing a gradient, warns that hill-climbing tends to deteriorate on larger problems, and takes aim at a small example from the PDP chapter: a network with 4 inputs and 4 hidden units that needed 45,200 training presentations, each of its 16 patterns shown 2,825 times, to learn parity. Is that a good or a bad result, the authors asked, without a theory of why so many trials are needed? Olazaran's thesis sets the two evaluations side by side as another case of the same results read in opposite ways.

Papert, writing in Daedalus the same year, asked in print whether he and Minsky had tried to kill connectionism and answered that there had been some hostility, as Quote Investigator also reproduces. The question had become part of the field's understanding of itself.

The honest limits

The book proved less than its reputation. Its theorems concern a specific architecture: fixed, local feature detectors feeding a single trainable threshold unit, with locality measured by order or by diameter. It says nothing rigorous about networks whose hidden units are themselves trained, which is where the field eventually went. The page proofs of chapter 12, circulated as MIT AI Memo 167 in October 1968, describe that chapter as written more to incite student research than to offer solutions, and the multilayer verdict was openly labeled intuitive. Treating a conjecture as though it were a theorem was the readers' error more than the book's.

The objections have limits too. Parity and connectedness genuinely are hard for local parallel devices, and Block's observation that people find them hard as well did not make the theorems any less true. The epilogue's warning that small successes need not scale was sound method in 1988, even though larger datasets and faster hardware later overtook its specific doubt about gradient training (see AlexNet). The analysis also had a technical afterlife at MIT: in 1976 Harold Abelson extended the order analysis to machines built from a fixed number of threshold units, concluding that such machines cannot recognize any topological invariant.

Finally, the counterfactual cannot be tested. No source can show what funders would have done without the book, and most of the participants' recollections were recorded in the late 1980s, after the revival, when every side had reasons to shade them.

How to place Perceptrons in AI history

Perceptrons sits between the 1958 perceptron and the funding contraction of the 1970s, and its most accurate label is the rigorous negative result that a field used to settle an argument it was already finishing. Its lasting technical contribution is the idea of measuring a parallel device by how much of its input each part must see. Its lasting historical lesson is about the gap between a proof and its reading: the theorems were about one architecture, while the reading was about an entire research program.

The authors' standing amplified that reading. Minsky received the A.M. Turing Award in 1969, the year the book appeared, according to MIT's obituary of Minsky. The MIT Press reissued the expanded edition in 2017 with a foreword by Bottou, and the publisher's description of that reissue quotes his verdict that the authors' rigorous work "does not make the perceptron look very good". Read today, the book is both a model of how to state precisely what a learning machine cannot do and a case study in how an intuitive judgment, printed beside real theorems, can travel further than the theorems themselves.

References

• Minsky, M. and Papert, S. (2017). Perceptrons: An Introduction to Computational Geometry, reissue of the 1988 expanded edition.
• Minsky, M. and Papert, S. (1967). Linearly Unrecognizable Patterns, MIT AI Memo 140.
• Minsky, M. and Papert, S. (1968). Linear Separation and Learning, MIT AI Memo 167.
• Newell, A. (1969). A Step toward the Understanding of Information Processes, review of Perceptrons.
• Rosenblatt, F. (1958). The perceptron: a probabilistic model for information storage and organization in the brain.
• Rosenblatt, F. (1961). Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms.
• Olazaran, M. (1991). Historical Sociology of Neural Network Research, PhD thesis, University of Edinburgh.
• Olazaran, M. (1996). A Sociological Study of the Official History of the Perceptrons Controversy.
• Nilsson, N. J. (2009). The Quest for Artificial Intelligence, chapter 16: Conferences, Books, and Funding.
• Rumelhart, D., Hinton, G. and Williams, R. (1986). Learning representations by back-propagating errors.
• Quote Investigator (2026). Quote Origin: Our Intuitive Judgment That the Extension Is Sterile.
• Bottou, L. (2015). Perceptrons Revisited, AAAI Spring Symposium talk slides.
• Abelson, H. (1976). Computational Geometry of Linear Threshold Functions, MIT AI Memo 376.
• Cornell University (2019). Professor's perceptron paved the way for AI, 60 years too soon.
• MIT News (2016). Professor Emeritus Seymour Papert, pioneer of constructionist learning, dies at 88.
• MIT News (2016). Marvin Minsky, father of artificial intelligence, dies at 88.