SNARC

Minsky and Edmonds built a net of probabilistic artificial neurons, of valves and potentiometers, that learned to run a maze.

The SNARC was built in the summer of 1951 at the Harvard Psychological Laboratory, under a US Air Force contract. In each unit a potentiometer set the probability of passing a pulse on, and the unit's internal memory set how long its clutch stayed engaged on the reinforcement shaft; the reinforcement process and the memory span ran on the order of half a second to ten seconds. A reward pulse turned up the potentiometers of units that had recently fired, a punishment pulse turned them down. The mazes had up to ten vertices. With wires pulled out and valves removed the net, after a period of confusion, adapted again to the new situation, as long as not too many units were put out of action. The thesis does not say how many units there were in its description of the machine (section 4/5.2, pp. 4-33 to 4-40). In the sources read, the number forty appears only in Minsky's own later accounts. In a 1981 interview for The New Yorker, as reproduced on the cyberneticzoo site, he says he thought the design would work well with about forty neurons and that the machine's memory sat in the positions of forty control knobs; in 2011 the machine is a rack of forty such units. The money, by the 1981 interview, is a couple of thousand dollars that George Miller got from the Office of Naval Research; by the 2011 account it came from the Air Force Office of Scientific Research, also through Miller. In the thesis the Office of Naval Research supports the thesis work itself in part, through a research project of Princeton's mathematics department. The proposal for the Dartmouth summer project, of 31 August 1955, says only, without naming the machine, that Minsky had built a machine for simulating learning by nerve nets; it gives no number of units and no date of building. What the record does not claim: that a document of 1951-52 giving the number of units or the date of building has been read. The report of the Harvard Psychological Laboratories titled A Neural-Analogue Calculator Based upon a Probability Model of Reinforcement, which Minsky's own list of publications dates 8 January 1952 and says describes the hardware of the SNARC, was not found in an open copy and has not been read: no copy of it turned up in Google, archive.org or HathiTrust, only mentions in other works, the Harvard library catalogue has no record of it, and DTIC's search was under maintenance when tried. It is not in the thesis bibliography.