Mark I Perceptron: the perceptron in hardware
In spring 1960 engineers at the Cornell Aeronautical Laboratory described the Mark I, Rosenblatt's perceptron in hardware. A retina of 20 x 20 photoresistors, 512 association units, 8 response units, and weights that were potentiometers turned by electric motors. After sixteen exposures to each letter the machine identified all 26 letters of a standard font.
Why it matters
Learning in a network got purpose-built hardware in which what had been learned was held in the positions of hundreds of potentiometers rather than in the memory of a general-purpose machine. Unlike SNARC in 1951, it was a machine built to test a theory and used for experiments in learning to recognise.
By Hay and Wightman's Research Trends article: the 400 sensory units have 16,000 possible connections on a patch board, and their links to the association units are scattered at random, so each association unit sees a random spread of points on the retina. Electromechanical parts were chosen for economy and expediency; the long time constant of the motor integrators, the authors write, could not be matched within practical bounds of availability and price. In an experiment with 30 per cent of the trainer's reinforcements deliberately wrong, both halves of the machine ended above the trainer's own accuracy. The research was sponsored by the Office of Naval Research from July 1957, with the Rome Air Development Center from 1958. In Research Trends in summer 1958 Rosenblatt wrote that a working model was scheduled for completion within the next year: the perceptron then existed as theory and as a simulation on an IBM 704. The Mark I operators' manual (Cornell Aeronautical Laboratory report VG-1196-G-5), dated 15 February 1960, a month before the IRE convention paper, describes the machine as an existing one and gives the same numbers: 400 photoresistors in a 20 x 20 array; 512 association units, eight to a printed card, each with its output from a potentiometer turned by a small D.C. motor, the wiper positions of all 512 being the memory, which moves about one-sixteenth of a revolution per minute at the full feedback voltage of 50 volts; eight response units with adjustable thresholds; and an IBM 704 program that generates the pseudo-random wiring between sensory and association units. Documents of the time place the first public demonstration of the Mark I in June 1960, at the Cornell Aeronautical Laboratory in Buffalo. The Digital Computer Newsletter of the Office of Naval Research (ONR), July 1960, says the machine "was demonstrated publicly for the first time in June" at the laboratory. The weekly electronics, in its issue dated 24 June 1960 (p. 43), says it was so demonstrated "this week", and quotes M. C. Yovitz of ONR and Warren Dunn of the Rome Air Development Center. Science News Letter of 2 July 1960 (p. 7) says it was demonstrated publicly for the first time in Buffalo, and Aviation Week of 4 July 1960 (p. 72, datelined Buffalo) says "recently". None of the four documents gives a day or names an audience beyond the officials quoted. By the ONR newsletter and electronics, letters of a single type face were shown at the demonstration and identified without error; by the newsletter and Science News Letter, with another type face added, the machine was right 79 per cent of the time. Among what was read, the day 23 June 1960, and the fact that it was a demonstration for the press, come only from the finding aid of the Charles Babbage Institute at the University of Minnesota (collection CBI 48), prepared in March 1991, and from its WorldCat record: a description of a folder of 1960 papers (prepared statements, attendee list, press conference agenda, clippings, short reports by Albert E. Murray, Frank Rosenblatt and Marshall C. Yovitz). It is a later catalogue description and does not say where the day comes from. "This week" in an issue dated Friday 24 June does not exclude Thursday 23 June, but does not name it. So the record keeps its date at March 1960, when the machine was described at the IRE convention; the public demonstration, by documents of the time, came later, in June. The number of exposures to each letter differs between documents of the time. Hay and Wightman's article in Research Trends in spring 1960, as read for this record, gives sixteen. Aviation Week (in a "representative demonstration") and Science News Letter give 15. The ONR newsletter says that for the 26-letter test, "which had been demonstrated", the machine reached an errorless level "after 40 exposures to each letter". The record keeps the sixteen of the engineers' article and does not decide which number is the more exact. What the record does not claim. The manual mentions neither 26 letters, nor sixteen exposures, nor an experiment on letters, nor a public demonstration, nor June 1960: it is dated 15 February 1960 and says how to operate the machine, not what it was shown to do, so it neither confirms nor refutes any date of a public demonstration. The documents of June and July 1960, now read, place the first public demonstration in June 1960 but give no day. The day, 23 June, stands in the Babbage Institute's finding aid and in the WorldCat catalogue, and later works repeat it; the folder of 1960 papers they describe (statements, attendee list, agenda, clippings) was not read. The record also does not claim that the machine first ran in June 1960: no document read says when it did, and the manual of 15 February 1960 describes it as an existing one. Nor does it claim that the numbers of exposures given above agree.