11 October 2026 · History · about 18 minutes
The Embryo and the Funeral: Rosenblatt's Perceptron, the Press That Celebrated It, and the Book That Buried It
Frank Rosenblatt told the Navy his machine would one day be conscious. Marvin Minsky had once built the same kind of machine. What happened next took a quarter-century and more to undo.
The Press Conference
In July 1958, the United States Office of Naval Research held a press conference in Washington at which the perceptron was described as "the first machine which is capable of having an original idea," in the words of its creator, Frank Rosenblatt.1 In objective terms, the machine had done something modest: after fifty trials with a deck of punch cards, an IBM 704 had taught itself to tell cards marked on the left from cards marked on the right. Rosenblatt was a research psychologist not yet thirty, and he did not describe what had happened in modest terms. The perceptron was not a fraud, and Rosenblatt was not a fool. But the next morning, the New York Times had assigned it a future it could not yet earn; by November the New Yorker had amplified the promise still further.
Rosenblatt had written: "Stories about the creation of machines having human qualities have long been a fascinating province in the realm of science fiction. Yet we are about to witness the birth of such a machine – a machine capable of perceiving, recognizing and identifying its surroundings without any human training or control."2 The Times published its report under the headline: "NEW NAVY DEVICE LEARNS BY DOING: Psychologist Shows Embryo of Computer Designed to Read and Grow Wiser."3 Based on Rosenblatt's statements at the press conference, the Times reported the perceptron to be "the embryo of an electronic computer that [the Navy] expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence."4 In late November, the New Yorker weighed in: "Indeed, it strikes us as the first serious rival to the human brain ever devised."5
This was a great deal to claim for a machine that had learned to sort punch cards. To understand what happened next, the excitement, the backlash, the years of stagnation, and the eventual vindication, it helps to begin earlier, with the ideas Rosenblatt had inherited and the problem he was genuinely trying to solve.
The Lineage
The story of the perceptron begins, as so much in early AI does, with a 1943 paper by Warren McCulloch and Walter Pitts. McCulloch was a neurophysiologist. Pitts was something stranger: he left home in Detroit at fifteen to study and went to Chicago.6 At the University of Chicago he sat in on Bertrand Russell's lectures without registering as a student. Russell was a visiting professor at Chicago in the fall of 1938, and he directed Pitts to go study with the logician Rudolf Carnap.7 Pitts met Carnap by walking into his office during office hours and presenting him with an annotated version of Carnap's recent book on logic, The Logical Syntax of Language. Since Pitts did not introduce himself, Carnap spent months searching for him; when he found him, he obtained a menial job at the university for Pitts, who was at that time homeless and without income.8 McCulloch later invited Pitts and his friend Jerome Lettvin to live with his family; their evening collaboration produced the paper that would follow.
Together they published "A Logical Calculus of the Ideas Immanent in Nervous Activity" in the Bulletin of Mathematical Biophysics, proposing the first mathematical model of the nervous system as a network of simple logical elements.9 Their artificial neurons received inputs, performed a weighted sum, and fired an output signal based on a threshold function; by connecting them in various configurations, McCulloch and Pitts showed their model could perform all logical functions.10
Six years later, Donald Hebb, a psychologist at McGill University, published The Organization of Behavior (1949), proposing a rule for how synapses learn. The connection between two neurons is strengthened when one repeatedly and persistently takes part in firing the other.11 The phrase later distilled from this, "cells that fire together, wire together," was not Hebb's own, but the idea was, and it supplied the engineering intuition that Rosenblatt would translate into mathematics.
The McCulloch-Pitts neuron was clever but rigid: its connections were fixed. What Rosenblatt added was the ability to learn. He introduced modifiable, continuously valued weights between units. The key innovation was the learning rule: modifiable weights that adjusted in response to error, absent from the fixed McCulloch-Pitts model.12
Marvin Minsky, who would later become Rosenblatt's most formidable critic, had in 1951 built one of the first hardware neural networks, a machine called SNARC, the Stochastic Neural Analog Reinforcement Calculator. Minsky was then a graduate student in mathematics at Princeton; his collaborator Dean Edmonds was a physics graduate student there who volunteered that he was good with electronics.13 SNARC comprised roughly forty Hebb synapses and modelled a rat running a maze, built partly from the surplus autopilot of a B-24 bomber.14 That the man who built one of the first hardware neural networks would later write the book credited with ending neural network research for a generation is one of the odder ironies in the history of computing.
The world into which Rosenblatt's perceptron arrived in 1958 was not empty. The Dartmouth workshop of 1956 had given the field its name and its founding ambitions. The Logic Theorist, written by Allen Newell, Herbert Simon, and Cliff Shaw and brought to wider attention around that time, could prove theorems from Principia Mathematica by manipulating symbols according to formal rules; it became the signature achievement of symbolic AI as a research programme.15 McCarthy, Minsky, Newell, and Simon were already working on programs that manipulated logical symbols rather than adjusting numerical weights. Their approach competed for the same limited federal funds, and the scale of that competition was not equal. The ONR research manager Marvin Denicoff later explained that ONR, rather than ARPA, had funded the perceptron project precisely because it was unlikely to produce technological results in the near or medium term: funding from ARPA ran to the order of millions of dollars, while funding from ONR was on the order of ten thousand dollars.16 J. C. R. Licklider, who had been interested in self-organising and biologically inspired methods in the 1950s, became head of IPTO at ARPA in 1962 already moving against them, and by the mid-1960s was openly critical of the perceptron; instead he strongly favoured the logical AI approach of Simon and Newell.17 The dispute between connectionist and symbolic AI was never purely a matter of mathematics. It was also a matter of which style of work the primary funder believed in, and the primary funder's money was never in Rosenblatt's corner. The Mansfield Amendment of 1969, which required DARPA to fund only research with a direct relationship to specific military functions, squeezed symbolic AI's open-ended research budget as well, and it is too simple to read the funding history as a story of one side picking winners.18
Rosenblatt was not the only person developing a learning network in this period. At Stanford in 1960, Bernard Widrow and his doctoral student Marcian Hoff developed ADALINE, the Adaptive Linear Neuron, a single-layer network trained by the least mean squares algorithm, which minimised the squared error on every training example rather than responding only to outright misclassification.19 ADALINE found commercial application in adaptive noise cancellation, including the elimination of echoes on long-distance telephone transmission lines; MADALINE, the multilayer version, was demonstrated on tasks including weather forecasting and speech recognition.20
The Boy from New Rochelle
Frank Rosenblatt was born in New Rochelle, New York, on 11 July 1928, the son of Dr. Frank and Katherine Rosenblatt.21 He attended the Bronx High School of Science, the same school as Minsky, a year apart from him, before going to Cornell, where he took his undergraduate degree in social psychology in 1950 and his doctorate in 1956.22 Even his dissertation showed what kind of mind he had: he built a custom computer, the Electronic Profile Analyzing Computer, to run multidimensional analysis on psychometric data from more than two hundred Cornell undergraduates; fifteen minutes of data processing took just two seconds.23
After his doctorate he moved to the Cornell Aeronautical Laboratory in Buffalo, New York, where he worked successively as a research psychologist, senior psychologist, and eventually head of the cognitive systems section.24 It was there, beginning in 1957, that he developed the perceptron. His first technical report, titled "The Perceptron: A Perceiving and Recognizing Automaton," appeared in January 1957 as a Cornell Aeronautical Laboratory report under Project PARA; the landmark journal paper, "The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain," was published in Psychological Review in 1958.25 The title is telling. He was not, in his own mind, building a machine. He was doing psychology. The perceptron was a model of the brain, a hypothesis about how biological neural systems might actually store and retrieve information.
His conception of the device had three layers. An array of sensory units detected the input. A hidden layer of association units processed signals from the sensory layer along connections that were set randomly at the outset: randomly, because Rosenblatt believed the retina was itself randomly connected to the visual cortex, and he wanted his machine to resemble human visual perception.26 The association units then fed response units whose weighted connections were the ones that actually learned: encoded in potentiometers, with weight updates performed by small electric motors.27 His 1958 paper opened with three questions about the brain's information-processing capabilities: how does it sense the physical world, in what form is information stored in memory, and how does information in memory influence recognition and behaviour?28 These are still the right questions.
Rosenblatt's laboratory culture rewarded boldness over caution. He once recruited a Cornell freshman named Terrell Koken over a conversation about Fermat's Last Theorem in the music lounge; Koken later recalled that Rosenblatt would arrive at a problem, deliver a verdict on it, and move on before anyone else had started.29 The Mark I Perceptron, assembled at the Cornell Aeronautical Laboratory in 1959 and first shown in public in June 1960, had an input retina of four hundred photocells arranged in a twenty-by-twenty grid, a hidden layer of 512 association units, and eight response units.30 His follow-up to the IBM 704 simulation, the Tobermory, named after the talking cat in an H. H. Munro story, was designed for speech recognition, had four layers and 12,000 weights implemented in toroidal magnetic cores, occupied an entire room, and was built between 1961 and 1967.31
The full scope of what Rosenblatt actually proved is easily lost in accounts that concentrate on the 1958 press conference and the 1969 rebuttal, and it is worth setting out before turning to the critique. His 1962 book Principles of Neurodynamics is the primary technical document of record. It proved that an elementary perceptron with an unlimited number of hidden-layer units could solve any classification problem, an existence theorem of genuine mathematical force.32 The book also studied cross-coupled perceptrons, machines with connections between units in the same layer, which are now recognised as precursors to Hopfield networks; back-coupled networks, with connections from later layers to earlier ones; four-layer perceptrons in which the last two layers had adjustable weights; and networks incorporating time delays for processing sequential data.33 When Jack Cowan recalled, in later years, that neural network research had already been slowing down by 1962 even before Minsky and Papert's intervention, part of the reason was that Principles of Neurodynamics itself was not widely absorbed by the field at the time of its publication; Minsky and Papert were responding primarily to the 1958 paper and the surrounding press coverage, not to Rosenblatt's most sophisticated work.34 The Minsky-Papert book published seven years later would address a far simpler subclass of machines than the 1962 book described.
The Hype Machine
The July 1958 press conference was arranged by the Office of Naval Research, which had funded the work. The Navy had its own reasons to emphasise the promise of a learning machine during the Cold War, and Rosenblatt had a temperamental preference for the large claim over the cautious one. His former doctoral student George Nagy later said: "On a scale of 1 to 10, he would have rated its significance 10."35
What is interesting, looking back, is how rational the excitement was at the level of the idea even if not at the level of the machine. The convergence theorem, proved by Rosenblatt, with H. D. Block providing a contemporaneous independent proof in his 1962 paper in Reviews of Modern Physics and Novikoff a third clean proof that same year, showed that if a classification problem could in principle be solved by a perceptron's architecture, then the learning algorithm would converge on a solution in a finite number of steps.36 This was a mathematically solid guarantee of learning. If the brain used something like a perceptron, and if the human brain could learn anything at all, then perhaps the perceptron could too. The leap from theorem to prediction was large, but it was not delusional. It was the kind of extrapolation that every field makes at the moment a new tool first arrives.
By 1959, Rosenblatt had moved to Cornell's main campus in Ithaca as director of the Cognitive Systems Research Program.37 The CIA's Photo Division studied the Mark I from 1960 to 1964 for the recognition of militarily interesting silhouetted targets, planes and ships, in aerial photographs; separately, a four-year programme at the National Photographic Interpretation Center ran from 1963 through 1966 to develop the algorithm into a working tool for photo-interpreters.38
Rosenblatt called Minsky "the loyal opposition."39 They had known each other since school; at conferences, Rosenblatt and Minsky publicly debated the perceptron's viability as their colleagues and students looked on. "Rosenblatt had a vision that he could make computers see and understand language," Thorsten Joachims, a machine learning professor at Cornell, has said. "And Marvin Minsky pointed out that that's not going to happen, because the functions are just too simple."40
The Loyal Opposition Writes a Book
When Seymour Papert arrived at MIT in 1963, he and Minsky decided to write a systematic mathematical account of the limitations of perceptrons.41 Papert brought a specific kind of mathematical training to the project. He held two doctoral degrees: the first from the University of the Witwatersrand in 1952, the second from the University of Cambridge, with a thesis titled The Lattices of Logic and Topology; sources differ on whether the Cambridge degree was awarded in 1959 or 1960, with the Mathematics Genealogy Project giving 1959 and the Cambridge thesis catalogue record giving 1960.42 He had also spent several years at the University of Geneva working with Jean Piaget on how children learn mathematics. It was Papert's topological and geometric apparatus that gave the critique of the perceptron the formal depth that Minsky's earlier informal objections had lacked. It took them until 1969 to finish, not because the project was simple but because mathematical problems kept unexpectedly requiring resolution as they wrote.43 Before the book appeared, an unpublished technical manuscript had already circulated in the field and was used, by some accounts, to redirect funding away from neural network research toward symbolic AI.44
Perceptrons: An Introduction to Computational Geometry, published by MIT Press in 1969, is a work of genuine mathematical elegance.45 Its central results concern parity and connectedness. Minsky and Papert took as their subject abstract versions of a class of learning devices, called perceptrons, "in recognition of the pioneer work of Frank Rosenblatt."46 These were single-layer machines in which each unit received connections from only a small part of the input, what the authors called conjunctive localness. Under this restriction, Minsky and Papert proved that a single-layer perceptron cannot compute the parity function without the order of connections growing without bound as the input grows, and cannot reliably determine whether a figure in the input retina is connected without connections growing at least as fast as the square root of the input size.47 These are genuine mathematical results. The book was dedicated to Rosenblatt.
H. D. Block was better placed than almost anyone to identify what Minsky and Papert had ignored. His 1970 review argued that they had chosen to study "a severely limited class of machines from a viewpoint quite alien to Rosenblatt's," and that the title of the book was "seriously misleading."48 Block's objection had a precise technical basis: Principles of Neurodynamics had proved that an unrestricted elementary perceptron with sufficiently many hidden-layer units can solve any classification problem; it had studied four-layer networks, back-coupled networks, and cross-coupled networks now known as Hopfield networks. Minsky and Papert had taken the simplified, single-layer, conjunctively local perceptron and written a book about it. The results were correct. The subject was narrow. The field treated the former as establishing the latter. Bernard Widrow complained that Minsky and Papert had defined perceptrons too narrowly, and said their proofs were "pretty much irrelevant," coming a full decade after Rosenblatt's perceptron.49
Allen Newell, reviewing the book for Science, opened his piece by declaring "this is a great book."50
The Problem With the Problem
In a 1971 Progress Report circulated to funding agencies under Minsky and Papert's names, they wrote of what they called Gamba perceptrons, two-layer machines in which the first layer is also made of linear threshold units rather than boolean units, that "virtually nothing is known about the computational capabilities of this latter kind of machine. We believe that it can do little more than can a low order perceptron."51 This claim was directed at funders at a moment when funding decisions were being made. It was also wrong.
Mikel Olazaran's sociological study of the perceptrons controversy, published in Social Studies of Science in 1996, argues that the pessimistic reading of the book was not a passive misreading by a confused community but something actively constructed and promoted. Researchers in symbolic AI instrumentalised the book to argue that neural network funding was wasteful; in Olazaran's account, this constituted the official history of the controversy, a story manufactured and circulated for strategic purposes, not merely a natural consequence of the book's technical claims.52 Minsky himself later compared Perceptrons to the Necronomicon in H. P. Lovecraft's fiction, describing it as often cited but rarely read.53
When the 1988 expanded edition of Perceptrons appeared, Minsky and Papert had a second opportunity, after nearly twenty years, to correct the misreading their book had suffered. They did not take it. The prologue and epilogue added to that edition engaged with the 1980s neural network revival, but Minsky and Papert largely doubled down, repeating their 1969 judgment that multilayer nets were a "sterile" extension and that Gamba machines would require "an enormous number" of Gamba-masks; the edition added no new scientific results of its own, only conjecture.54 The 1988 edition is the strongest evidence in the whole episode that the pessimism was not merely a misreading by others but a position Minsky and Papert actively maintained even as experiments were refuting it. These conjectures were wrong.
The algorithm that would prove them wrong had been sitting in a Harvard doctoral dissertation since 1974. Paul Werbos had described backpropagation, the technique of training multilayer networks by propagating error signals backward through the layers, as part of his thesis, well before the field took notice.55 Rumelhart developed it independently in spring 1982; his 1985 paper did not cite Werbos's earlier work, as he and his collaborators did not know of it.56 When Rumelhart, Geoffrey Hinton, and Ronald Williams published their landmark paper in Nature in 1986, they claimed to have overcome the problems Minsky and Papert had identified, arguing that "their pessimism about learning in multilayer machines was misplaced."57 That same year, Rumelhart and James McClelland published Parallel Distributed Processing, a two-volume work whose subtitle, Explorations in the Microstructure of Cognition, announced an ambition to reconnect neural network modelling with cognitive science as a whole; together with the Nature paper it turned the connectionist revival from a technical result into a movement.58
The End and the Aftermath
Rosenblatt did not live to see the consequences of Perceptrons fully unfold. He died on 11 July 1971, his forty-third birthday, while sailing a sloop called the Shearwater in Chesapeake Bay.59 He was eulogised in the Congressional Record by Representative Hugh Carey of New York, and separately by former Senator Eugene McCarthy, for whose 1968 presidential campaign Rosenblatt had worked in New Hampshire and California.60 When he died, his research had moved far from the perceptron: from 1966 onward he had become fascinated with whether learned behaviour could be transferred from trained to untrained rats by the injection of brain extracts. His own later results showed that the initial large effects were probably not real, and that any memory transfer was at most very small, a willingness to follow evidence to an uncomfortable destination that was characteristic of the man.61
Richard O'Brien, former director of Cornell's Division of Biological Sciences, said at Rosenblatt's memorial service: "He would never, as prudent professors do, pick out some small aspect of the physical universe which could be studied and could reasonably be expected to produce solutions in a year or two. Instead, he would reach out and grasp the biggest problems that he could see, and apply himself, throw himself into the study of them, without any recognition of the fact that if he made bad choices, or if he chose, as he usually did, problems which were not likely to yield a solution within 10 or 20 years, that this would redound to his disadvantage. He never worked that way."62
Minsky received the ACM Turing Award for 1969, computing's highest honour.63 Neural network research did not entirely disappear in the 1970s. A few researchers, James Anderson, Teuvo Kohonen, Stephen Grossberg, continued to work on connectionist systems, though they were very much the exception rather than the rule. The AI funding that had been flowing toward perceptrons moved to symbolic AI. It is worth being precise about what happened next, since two different stories are easily conflated. The decline of neural network research in the decade after 1969 was attributable in part to the way Perceptrons was deployed, in part to Licklider's prior turn against biologically inspired methods at IPTO, and in part to broader forces that affected all of AI. The Mansfield Amendment of 1969 squeezed DARPA's open-ended research budget without regard to research style, and the UK Science Research Council's Lighthill Report of 1973 concluded that AI had failed to deliver on its central claims, generating consequences for British funding that fell on symbolic and connectionist work alike. The broader AI winter was a separate event from the neural-network-specific contraction: it came from the failure of symbolic AI's own promises, and it arrived for different reasons at different times on different sides of the Atlantic.
The revival when it came had two distinct lines. In 1982, John Hopfield published a paper on content-addressable memory: he showed that a recurrent network of simple threshold units would settle into stable states that could store and retrieve patterns, and he described this behaviour using an energy function borrowed from statistical physics.64 Hopfield's recurrent networks were the direct descendants of what Rosenblatt had called close-loop cross-coupled perceptrons in Principles of Neurodynamics, though the two lines of work arrived at similar architectures by different routes.65 Separately, Rumelhart, Hinton, and Williams made backpropagation practical for multilayer feedforward networks in their 1986 Nature paper.66 Then in 1989, Yann LeCun and colleagues at AT&T Bell Laboratories published a paper in Neural Computation demonstrating backpropagation applied to a convolutional network for recognising handwritten zip-code digits, the direct heir, in vision terms, to what Rosenblatt had been trying to do with the Mark I Perceptron three decades earlier.67
The Mark I Perceptron now resides in the Smithsonian Institution in Washington, DC. The IEEE established the Frank Rosenblatt Award in 2004.68 "What Rosenblatt wanted," Joachims has said, "was to show the machine objects and have it recognize those objects. And 60 years later, that's what we're finally able to do. So he was heading on the right track, he just needed to do it a million times over."69
What to Make of It
Frank Rosenblatt drowned at forty-three on his birthday, sailing into a bay on the same day that, thirteen years earlier, a navy press officer had told the world his machine would one day be conscious. The vindication he did not live to see arrived in exactly the form he had predicted, through vision systems trained by iterative adjustment of numerical weights, and came from researchers who had to rediscover much of what he had already proved.
What the episode illustrates, and what is worth carrying away, is the difference between a mathematical result and its reception. Rosenblatt's existence theorem was correct in 1962; the convergence theorem was correct in 1958; neither fact protected his programme from being read into irrelevance by a book that proved something narrower and truer, and that was then deployed, as Olazaran's sociological study documents, to serve the interests of a competing research community that controlled far more money. Rumelhart did not know of Werbos's work when he reinvented backpropagation; Hopfield's recurrent networks turned out to be what Rosenblatt had sketched twenty years earlier. The gap between a proof and the story told about a proof is where most of the damage in this episode was done, and it is a gap that can open at any moment in the history of science, whenever the politics of funding and the sociology of citation diverge from the underlying mathematics.
Notes
- Press conference date of 7 July 1958; Rosenblatt's claim that the perceptron was "the first machine which is capable of having an original idea": Melanie Lefkowitz, "Professor's perceptron paved the way for AI – 60 years too soon," Cornell Chronicle, 25 September 20191.
- Rosenblatt's 1958 written statement: Lefkowitz, Cornell Chronicle, 20191.
- New York Times headline, 8 July 1958: Lefkowitz, Cornell Chronicle, 20191.
- "the embryo of an electronic computer that [the Navy] expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence": Perceptron, Wikipedia2, citing Olazaran 1996.
- New Yorker, "Rival," late November 1958: Lefkowitz, Cornell Chronicle, 20191; bibliographic details: Mason, Harding; Stewart, D.; Gill, Brendan (28 November 1958), "Rival," The New Yorker: Frank Rosenblatt, Wikipedia3.
- Pitts born in Detroit; left home at fifteen to go to Chicago: Walter Pitts, Wikipedia4.
- Russell as a visiting professor at Chicago in the fall of 1938, directing Pitts to study with Carnap: Walter Pitts, Wikipedia4.
- Pitts presenting annotated copy of The Logical Syntax of Language to Carnap; Carnap spending months searching for him; obtaining a menial job for Pitts, who was homeless and without income: Walter Pitts, Wikipedia4 ("Carnap spent months searching for him, and, when he found him, he obtained for him a menial job at the university and had Pitts study with him. Pitts at the time was homeless and without income").
- McCulloch and Pitts, "A Logical Calculus of the Ideas Immanent in Nervous Activity," Bulletin of Mathematical Biophysics 5 (1943): 115–133; McCulloch inviting Pitts and Lettvin to live with his family: Walter Pitts, Wikipedia4.
- McCulloch-Pitts neurons performing logical functions through threshold networks: "A Logical Calculus of the Ideas Immanent in Nervous Activity," Wikipedia5.
- Hebb's 1949 learning postulate: Donald O. Hebb, The Organization of Behavior (Wiley, 1949): "Donald Hebb Formulates the Hebb Synapse," History of Information6.
- Rosenblatt's modifiable continuously valued connections (weights) adjusting in response to error: I. S. N. Berkeley, "A Revisionist History of Connectionism," University of Louisiana7.
- SNARC: Minsky as a mathematics graduate student at Princeton; Edmonds as a physics graduate student volunteering his electronics skills: Stochastic Neural Analog Reinforcement Calculator, Wikipedia8.
- SNARC built partly from the autopilot of a B-24 bomber; approximately 40 Hebb synapses: "SNARC — the first neural network in hardware," History of AI9; SNARC, Wikipedia8.
- The Logic Theorist, written by Newell, Simon, and Cliff Shaw, capable of proving theorems from Principia Mathematica by symbolic manipulation: Dartmouth workshop, Wikipedia10.
- ONR research manager Marvin Denicoff explaining that ONR rather than ARPA funded the perceptron; "Funding from ARPA go up to the order of millions dollars, while from ONR are on the order of 10,000 dollars": Perceptron, Wikipedia2, citing Penn 2021.
- Licklider interested in biologically inspired methods in the 1950s; became IPTO head in 1962; by the mid-1960s openly critical of the perceptron; strongly in favour of the logical AI approach of Simon and Newell: Perceptron, Wikipedia2, citing Guice 1998.
- The Mansfield Amendment of 1969, legislated as part of the Defense Authorization Act of 1970, forbidding military funding for any research without "a direct or apparent relationship to a specific military function or operations," and its effects on DARPA and NSF computing programmes: National Research Council, Funding a Revolution: Government Support for Computing Research, chapter 4, National Academies Press, 199911.
- ADALINE developed by Widrow and Hoff at Stanford in 1960, using the least mean squares algorithm: ADALINE, Wikipedia12.
- ADALINE and adaptive echo cancellation on long-distance telephone transmission lines: Bernard Widrow and Michael A. Lehr, "30 Years of Adaptive Neural Networks: Perceptron, Madaline, and Backpropagation," Proceedings of the IEEE 78(9), 199013, confirming "adaptive echo cancellers for long-distance telephone and satellite circuits"; MADALINE on weather forecasting and speech recognition: ADALINE, Wikipedia12.
- Rosenblatt's birth, family, New Rochelle: Frank Rosenblatt, Wikipedia3.
- Rosenblatt at the Bronx High School of Science; Minsky a year apart; Cornell degrees: Lefkowitz, Cornell Chronicle, 20191; Frank Rosenblatt, Wikipedia3.
- Rosenblatt's Electronic Profile Analyzing Computer; data from more than 200 Cornell undergraduates; processing time of two seconds: Frank Rosenblatt, Wikipedia3.
- Rosenblatt at Cornell Aeronautical Laboratory, successively research psychologist, senior psychologist, and head of the cognitive systems section: Frank Rosenblatt, Wikipedia3.
- Rosenblatt's 1957 technical report titled "The Perceptron: A Perceiving and Recognizing Automaton" (Project PARA, Cornell Aeronautical Laboratory Report No. 85–460–1): Perceptrons (book), Wikipedia14. The 1958 paper: Rosenblatt, Frank (1958), "The perceptron: A probabilistic model for information storage and organization in the brain," Psychological Review 65(6): 386–408: Perceptron, Wikipedia2.
- Rosenblatt's random connections based on his belief about the retina-to-visual-cortex wiring: Perceptron, Wikipedia2.
- Adjustable weights encoded in potentiometers, updated by electric motors: Perceptron, Wikipedia2.
- Rosenblatt's three questions about the brain in his 1958 paper: "Kernelled Connections: Perceptron as Diagram," TripleAmpersand Journal15.
- Rosenblatt's recruitment of Terrell Koken over a Fermat's Last Theorem conversation: Lefkowitz, Cornell Chronicle, 20191.
- Mark I Perceptron architecture: 400 photocells, 512 A-units, 8 R-units; assembled at Cornell Aeronautical Laboratory between June 1959 and December 1959: Perceptron, Wikipedia2.
- Tobermory: four layers, 12,000 weights in toroidal magnetic cores, designed for speech recognition, built between 1961 and 1967, occupied an entire room: Frank Rosenblatt, Wikipedia3; Perceptron, Wikipedia2.
- Rosenblatt's existence theorem in Principles of Neurodynamics (1962): an elementary perceptron with unlimited hidden-layer units can solve any classification problem: Perceptrons (book), Wikipedia14; Frank Rosenblatt, Wikipedia3.
- Principles of Neurodynamics (Spartan Books, Washington DC, 1962): four-layer perceptrons, cross-coupled perceptrons (now known as Hopfield networks), back-coupled networks, time-delay networks: Perceptron, Wikipedia2; Frank Rosenblatt, Wikipedia3.
- Jack Cowan's recollection that neural network research had already been slowing by 1962; the field's limited engagement with Principles of Neurodynamics on publication: "Kernelled Connections: Perceptron as Diagram," TripleAmpersand Journal15 ("Jack Cowan recalls that even by 1962, research into neural networks had already been slowing down, and that this was only partially due to Minsky and Papert's criticisms").
- George Nagy on Rosenblatt's sense of the perceptron's significance: Lefkowitz, Cornell Chronicle, 20191.
- The convergence theorem proved by Rosenblatt; H. D. Block's contemporaneous independent proof in "The Perceptron: A Model for Brain Functioning," Reviews of Modern Physics 34:1 (1962): 123; the proof now standard is due to Novikoff (1962): Perceptron, Wikipedia2.
- Rosenblatt moving to Cornell's Ithaca campus in 1959 as director of the Cognitive Systems Research Program: Frank Rosenblatt, Wikipedia3.
- CIA Photo Division studying the Mark I from 1960 to 1964; NPIC programme from 1963 through 1966: Perceptron, Wikipedia2, citing O'Connor 2022.
- Rosenblatt calling Minsky "the loyal opposition": Lefkowitz, Cornell Chronicle, 20191.
- Thorsten Joachims on Rosenblatt's vision and Minsky's objection: Lefkowitz, Cornell Chronicle, 20191.
- Papert arriving at MIT in 1963; Minsky and Papert deciding to write a theoretical account of perceptron limitations: Perceptrons (book), Wikipedia14.
- Papert's first doctorate from University of the Witwatersrand (1952); second doctorate from Cambridge, thesis titled The Lattices of Logic and Topology: Mathematics Genealogy Project, Seymour Papert16 (giving 1959 as the year); the Cambridge thesis catalogue entry reads "Papert, Seymour Aubrey (1960). The lattices of logic and topology (PhD thesis). University of Cambridge" as recorded in Seymour Papert, Wikipedia17. Papert's work with Piaget at Geneva from roughly 1958 to 1963: "The South African mathematician who went to MIT," MyBroadband18.
- Minsky and Papert taking until 1969 because of unexpected mathematical problems that turned up as they wrote: Perceptrons (book), Wikipedia14.
- Prior unpublished manuscript circulated before the book, used to redirect funding: Berkeley, "A Revisionist History of Connectionism," citing Hecht-Nielsen 19907.
- Publication details of Perceptrons: An Introduction to Computational Geometry (MIT Press, 1969): Perceptrons (book), Wikipedia14.
- Minsky and Papert taking their subject "in recognition of the pioneer work of Frank Rosenblatt": Perceptrons (book), Wikipedia14.
- The parity theorem (Theorem 3.1.1) and the connectedness result (Section 5.5), both under conjunctive localness: Perceptrons (book), Wikipedia14.
- H. D. Block's review: Block, H. D. (1970), "A Review of 'Perceptrons: An Introduction to Computational Geometry,'" Information and Control 17: 501–522, arguing Minsky and Papert "study a severely limited class of machines from a viewpoint quite alien to Rosenblatt's" and that the title was "seriously misleading": Perceptrons (book), Wikipedia14.
- Bernard Widrow on Minsky and Papert defining perceptrons too narrowly; his comment that their proofs were "pretty much irrelevant": Perceptrons (book), Wikipedia14, citing Olazaran 1996, p. 630.
- Allen Newell's review opening with "this is a great book": Newell, Allen (1969), "A Step toward the Understanding of Information Processes," Science 165(3895): 780–782: Perceptrons (book), Wikipedia14.
- Minsky and Papert on Gamba networks: "Virtually nothing is known about the computational capabilities of this latter kind of machine. We believe that it can do little more than can a low order perceptron": Marvin Minsky and Seymour Papert, "Progress Report on Artificial Intelligence," MIT AI Memo AIM-252, 197119, as cited in Perceptrons (book), Wikipedia14.
- Olazaran's sociological study arguing that the pessimistic reading was actively constructed and promoted, constituting an "official history" of the controversy: Olazaran, Mikel (1996), "A Sociological Study of the Official History of the Perceptrons Controversy," Social Studies of Science 26(3): 611–659: cited and described in Perceptrons (book), Wikipedia14.
- Minsky comparing Perceptrons to the Necronomicon as often cited but rarely read: Berkeley, "A Revisionist History of Connectionism," citing Minsky, personal communication, 19947; confirmed as characterisation in Perceptrons (book), Wikipedia14.
- The 1988 expanded edition adding a prologue and epilogue but no new scientific results; conjecturing that multilayer nets are a "sterile" extension and that Gamba machines would require "an enormous number" of Gamba-masks: Perceptrons (book), Wikipedia14 ("An 'expanded edition' was published in 1988, which adds a prologue and an epilogue to discuss the revival of neural networks in the 1980s, but no new scientific results").
- Backpropagation first described by Paul Werbos in his 1974 Harvard doctoral dissertation: Paul Werbos, Wikipedia20.
- Rumelhart developed backpropagation independently in spring 1982; "The 1985 paper does not cite earlier publications of backpropagation, such as the 1974 dissertation of Paul Werbos, as they did not know the earlier publications": David Rumelhart, Wikipedia21.
- Rumelhart, Hinton, and Williams 1986 Nature paper claiming Minsky-Papert pessimism about multilayer machines "was misplaced": Perceptrons (book), Wikipedia14, citing Olazaran 1996; paper: Rumelhart, D.E., Hinton, G.E., Williams, R.J. (1986), "Learning representations by back-propagating errors," Nature 323: 533–536.
- Rumelhart and McClelland, Parallel Distributed Processing: Explorations in the Microstructure of Cognition, 2 vols. (MIT Press, 1986): MIT Press catalogue22; described as presenting "a new theory of cognition called connectionism that is challenging the idea of symbolic computation": Perceptrons (book), Wikipedia14.
- Rosenblatt's death on his forty-third birthday, 11 July 1971, while sailing a sloop called the Shearwater in Chesapeake Bay: Lefkowitz, Cornell Chronicle, 20191 ("In 1971, he drowned while sailing a sloop called the Shearwater in Chesapeake Bay on his 43rd birthday").
- The Congressional Record tribute entered by Representative Hugh Carey of New York: Frank Rosenblatt, Wikipedia3, citing Carey, Hugh L. (1971), "Tribute to Dr. Frank Rosenblatt," Congressional Record: Proceedings and Debates of the 92d Congress. Hugh Carey was a member of the U.S. House of Representatives for New York's 15th district from 1963 to 1974: Hugh Carey, Wikipedia23. Rosenblatt's work for Eugene McCarthy's 1968 presidential campaign in New Hampshire and California: Frank Rosenblatt, Wikipedia3.
- Rosenblatt becoming fascinated from 1966 with the transfer of learned behaviour from trained to naive rats by injection of brain extracts; his own later results showing that "the initial reports of larger effects were wrong and that any memory transfer was at most very small": Frank Rosenblatt, Wikipedia3.
- Richard O'Brien's memorial remarks: Lefkowitz, Cornell Chronicle, 20191.
- Minsky received the ACM Turing Award for 1969: ACM A. M. Turing Award, Marvin Minsky24.
- Hopfield's 1982 paper, "Neural networks and physical systems with emergent collective computational abilities," Proceedings of the National Academy of Sciences 79(8): 2554–2558: Hopfield network, Wikipedia25.
- Frank Rosenblatt studied "close-loop cross-coupled perceptrons", which are 3-layered perceptron networks whose middle layer contains recurrent connections that change by a Hebbian learning rule, recognised as precursors to Hopfield networks: Hopfield network, Wikipedia25.
- Rumelhart, Hinton, and Williams 1986 Nature paper making backpropagation practical for multilayer feedforward networks: David Rumelhart, Wikipedia21.
- LeCun, Boser, Denker et al. (1989), "Backpropagation Applied to Handwritten Zip Code Recognition," Neural Computation 1(4): 541–551: MIT Press, Neural Computation26; confirmed by LeNet, Wikipedia27.
- Mark I Perceptron at the Smithsonian; IEEE Frank Rosenblatt Award established 2004: Lefkowitz, Cornell Chronicle, 20191; IEEE Frank Rosenblatt Award, Wikipedia28.
- Thorsten Joachims on Rosenblatt heading in the right direction: Lefkowitz, Cornell Chronicle, 20191.